Merge remote-tracking branch 'origin/main'

.github/scripts/create-release.sh

@@ -38,7 +38,7 @@ if [ -z "${RELEASE_ID}" ]; then
 fi
 echo "Release created with ID ${RELEASE_ID}"
 
-for asset in dist/snapshot-*.whl dist/snapshot-*.tar.gz; do
+for asset in dist/*.whl dist/*.tar.gz; do
   if [ -f "${asset}" ]; then
     echo "Uploading ${asset}..."
     curl -fsS -X POST "${API_URL}/repos/${REPOSITORY}/releases/${RELEASE_ID}/assets" \

.github/workflows/build.yml

@@ -7,6 +7,8 @@ on:
     tags:
       - 'v*'
   pull_request:
+  
+  workflow_dispatch:
 
 jobs:
   ci:

design.md

@@ -62,7 +62,7 @@
 
 ## 🔌 3. 接口契约设计 (API Contract)
 
-### 接口：`POST /opencode-bridge/edit`（v2 — 无 `mode`）
+### 接口：`POST /opencode-bridge/edit`（v4 — 异步 + SSE）
 
 **Request Payload**（前端 `OpenCodeCellActions` 在用户提交 inline 输入框时发出）：
 
@@ -84,20 +84,40 @@
 }
 ```
 
-> **v2 变化**：body 不再含 `mode` / `action` 字段。前端只把 cell 的 input/output/error 收集到 `context`，加用户的自然语言 `prompt` 一起发出；后端用统一的 `UNIFIED_SYSTEM_PROMPT` + LLM 自己判断优化/排错/编辑。`providerId` / `modelId` 来自 JupyterLab settings 透传。
+> body 不含 `mode` / `action`。`providerId` / `modelId` 来自 inline 输入框动态选择，无 settings 持久化。
 
-**Response**（当前 v2 实现：单次 JSON；SSE 流式属于 Slice 4 后续工作）：
+**Response**（v4 异步）：后端调用 OpenCode `/session/:id/prompt_async` 后**立即返回**——不等 LLM 完成。
 
 ```json
 {
   "ok": true,
-  "finalSource": "import plotly.express as px\nfig = px.line(x=[1, 2, 3], y=[1, 2, 3])\nfig.show()",
   "sessionId": "ses-abc123",
   "notebookPath": "analysis/demo.ipynb"
 }
 ```
 
-> 成功时前端用 `finalSource` 就地替换 cell 源码（`cell.model.sharedModel.setSource`）。Session 由服务端 `SessionManager` 按 `notebookPath` 复用（1 notebook 1 sid），404 / `session not found` 时服务端自动 invalidate 并在下次请求重建。
+> 响应中**无** `markdown` 字段。AI 的回复通过独立的 SSE 端点（`/opencode-bridge/events`）实时推送给前端，前端在 inline prompt 的 history 区里增量渲染文本流 / 推理 / 工具调用 / 权限提问等。Session 仍由 `SessionManager` 按 `notebookPath` 复用（1 notebook 1 sid），404 / `session not found` 时服务端自动 invalidate 并在下次请求重建。cell 源码**不**被替换；用户可对 AI 消息中的每个代码块使用 Copy / Insert / Replace 按钮。
+
+### 接口：`GET /opencode-bridge/events?session=<sid>`（v4 新增 — SSE 代理）
+
+服务端长连接到 OpenCode Serve 的 `GET /global/event`，按 `session=<sid>` 过滤后以 `text/event-stream` 透传给前端。OpenCode 事件 schema 较松散，事件示例：
+
+```json
+{"type": "session.next.text.delta",        "properties": {"sessionID": "ses-abc123", "delta": "hel"}}
+{"type": "session.next.text.delta",        "properties": {"sessionID": "ses-abc123", "delta": "lo"}}
+{"type": "session.next.reasoning.delta",   "properties": {"sessionID": "ses-abc123", "delta": "thinking..."}}
+{"type": "session.next.tool.called",       "properties": {"sessionID": "ses-abc123", "name": "bash"}}
+{"type": "session.next.tool.input.delta",  "properties": {"sessionID": "ses-abc123", "delta": "{\"cmd\": \"ls\"}"}}
+{"type": "session.next.tool.success",      "properties": {"sessionID": "ses-abc123", "result": {"exitCode": 0}}}
+{"type": "session.next.agent.switched",    "properties": {"sessionID": "ses-abc123", "agentID": "build"}}
+{"type": "session.next.model.switched",    "properties": {"sessionID": "ses-abc123", "modelID": "claude-..."}}
+{"type": "file.edited",                    "properties": {"sessionID": "ses-abc123", "path": "foo.py"}}
+{"type": "permission.asked",               "properties": {"sessionID": "ses-abc123", "id": "perm-1", "description": "rm -rf"}}
+{"type": "question.asked",                 "properties": {"sessionID": "ses-abc123", "id": "q-1", "question": "Which env?"}}
+{"type": "session.idle",                   "properties": {"sessionID": "ses-abc123"}}
+```
+
+> v4 前端 `OpenCodeInlinePrompt.applyEvent()` 按 `type` 路由到不同渲染器：text/reasoning/tool 块用 `<details>` 可折叠，permission/question 提示独立块（响应需要 OpenCode Serve 提供对应 API；本扩展 v4 仅渲染，不实现 allow/answer 回调路由）。`session.idle` 触发流指针重置、关闭 SSE 订阅、重新启用输入框。System 事件（`server.*` / `workspace.*` / `lsp.*` / `mcp.*` / `pty.*` / `installation.*`）以日志行形式展示，不参与 session 过滤。
 
 ---
 

opencode_bridge/opencode_client.py

@@ -1,8 +1,22 @@
-"""Async HTTP client for the local OpenCode server."""
+"""Async HTTP client for the local OpenCode server.
 
+Two interaction modes:
+
+- **Request/response** (one-shot) - ``health``, ``list_providers``,
+  ``create_session``, ``send_message_async``, ``abort``, ``delete_session``,
+  ``list_session_messages``. ``send_message_async`` fires the prompt and
+  returns immediately; the actual LLM response is consumed via the SSE
+  stream (next mode).
+- **Streaming** (long-lived) - :meth:`stream_global_events` is an async
+  generator over OpenCode Serve's ``GET /global/event`` SSE feed. It
+  yields each event as a parsed JSON dict until the upstream closes.
+"""
+
+import asyncio
 import atexit
 import json
-from typing import Any, Optional
+import logging
+from typing import Any, AsyncIterator, Optional
 
 from tornado.httpclient import AsyncHTTPClient, HTTPRequest
 from tornado.httputil import HTTPHeaders
@@ -10,6 +24,9 @@ from tornado.httputil import HTTPHeaders
 from .config import OpenCodeConfig
 
 
+log = logging.getLogger("opencode_bridge.opencode_client")
+
+
 class OpenCodeError(Exception):
     """Raised when an OpenCode API request fails."""
 
@@ -64,22 +81,28 @@ class OpenCodeClient:
     async def create_session(self, title: str) -> dict[str, Any]:
         return await self._request("POST", "/session", {"title": title})
 
-    async def send_message_sync(
+    async def send_message_async(
         self,
         session_id: str,
         parts: list[dict[str, Any]],
         provider_id: Optional[str] = None,
         model_id: Optional[str] = None,
         system: Optional[str] = None,
-    ) -> dict[str, Any]:
+    ) -> None:
+        """Fire-and-forget POST /session/:id/prompt_async.
+
+        Returns once OpenCode has accepted the prompt and queued it for
+        processing - NOT when the LLM is done. Subscribe to
+        :meth:`stream_global_events` to follow the agent's progress.
+        """
         body: dict[str, Any] = {"parts": parts}
         if provider_id is not None and model_id is not None:
             body["model"] = {"providerID": provider_id, "modelID": model_id}
         if system is not None:
             body["system"] = system
-        return await self._request(
+        await self._request(
             "POST",
-            "/session/%s/message" % session_id,
+            "/session/%s/prompt_async" % session_id,
             body,
         )
 
@@ -87,6 +110,41 @@ class OpenCodeClient:
         result = await self._request("POST", "/session/%s/abort" % session_id)
         return result is not None
 
+    async def reply_permission(
+        self,
+        session_id: str,
+        permission_id: str,
+        response: str,
+    ) -> bool:
+        """Respond to a `permission.asked` event.
+
+        ``response`` must be one of ``"once"`` / ``"always"`` / ``"reject"``,
+        matching the OpenCode Serve API. Returns True on success.
+        """
+        result = await self._request(
+            "POST",
+            "/session/%s/permissions/%s" % (session_id, permission_id),
+            {"response": response},
+        )
+        return result is not None
+
+    async def reply_question(
+        self,
+        session_id: str,
+        question_id: str,
+        answer: str,
+    ) -> bool:
+        """Reply to a `question.asked` event with the user's freeform answer.
+
+        Returns True on success.
+        """
+        result = await self._request(
+            "POST",
+            "/session/%s/question/%s/reply" % (session_id, question_id),
+            {"answer": answer},
+        )
+        return result is not None
+
     async def delete_session(self, session_id: str) -> bool:
         result = await self._request("DELETE", "/session/%s" % session_id)
         return result is not None
@@ -95,9 +153,9 @@ class OpenCodeClient:
         """List all messages in the given OpenCode session.
 
         Returns the raw OpenCode response: a list of
-        `{ info: { role: "user"|"assistant", ... }, parts: [...] }`.
+        ``{ info: { role: "user"|"assistant", ... }, parts: [...] }``.
         The server extension is responsible for projecting this into a
-        frontend-friendly `{role, content}[]` shape.
+        frontend-friendly ``{role, content}[]`` shape.
         """
         return await self._request("GET", "/session/%s/message" % session_id)
 
@@ -105,6 +163,104 @@ class OpenCodeClient:
     def endpoint(self) -> str:
         return self._config.url
 
+    async def stream_global_events(self) -> AsyncIterator[dict[str, Any]]:
+        """Stream OpenCode Serve's ``GET /global/event`` SSE feed.
+
+        Yields one parsed JSON dict per ``data: {...}`` event block. The
+        stream runs until:
+
+        - the upstream disconnects (end of stream),
+        - the consumer stops iterating (caller cancels this generator),
+        - or a network error occurs (logged and the generator returns).
+
+        The consumer is responsible for filtering events by session ID
+        (the upstream pushes events for ALL sessions, plus system events).
+        """
+        queue: asyncio.Queue = asyncio.Queue()
+        end_sentinel = object()
+
+        def on_chunk(chunk: bytes) -> None:
+            if chunk:
+                queue.put_nowait(chunk)
+
+        headers = HTTPHeaders({
+            "Accept": "text/event-stream",
+            "Cache-Control": "no-cache",
+        })
+        request_kwargs: dict[str, Any] = {
+            "method": "GET",
+            "headers": headers,
+            "request_timeout": 0,  # 0 = no timeout for long-lived streams
+            # streaming_callback must live on the HTTPRequest itself; tornado
+            # rejects any fetch() kwargs when an HTTPRequest is passed.
+            "streaming_callback": on_chunk,
+        }
+        if self._config.auth is not None:
+            request_kwargs["auth_username"] = self._config.auth[0]
+            request_kwargs["auth_password"] = self._config.auth[1]
+
+        request = HTTPRequest(self._config.url + "/global/event", **request_kwargs)
+
+        async def _run_fetch() -> None:
+            try:
+                await self._http_client.fetch(request, raise_error=False)
+            except Exception as e:
+                log.warning("OpenCode /global/event fetch error: %s", e)
+            finally:
+                queue.put_nowait(end_sentinel)
+
+        runner_task = asyncio.create_task(_run_fetch())
+
+        buffer = ""
+        try:
+            while True:
+                chunk = await queue.get()
+                if chunk is end_sentinel:
+                    break
+                if not isinstance(chunk, bytes):
+                    continue
+                buffer += chunk.decode("utf-8", errors="replace")
+                while "\n\n" in buffer:
+                    block, buffer = buffer.split("\n\n", 1)
+                    parsed = self._parse_sse_block(block)
+                    if parsed is not None:
+                        yield parsed
+            # Drain any trailing data after the last \n\n
+            if buffer.strip():
+                parsed = self._parse_sse_block(buffer)
+                if parsed is not None:
+                    yield parsed
+        finally:
+            if not runner_task.done():
+                runner_task.cancel()
+                try:
+                    await runner_task
+                except (asyncio.CancelledError, Exception):
+                    pass
+
+    @staticmethod
+    def _parse_sse_block(block: str) -> Optional[dict[str, Any]]:
+        """Parse a single SSE event block into a JSON dict (or None).
+
+        SSE format reminder: a block is one or more ``field: value`` lines
+        separated by newlines; multiple ``data:`` lines are concatenated
+        with ``\\n``. We only care about ``data:``.
+        """
+        data_parts: list[str] = []
+        for line in block.split("\n"):
+            if line.startswith("data:"):
+                data_parts.append(line[5:].lstrip())
+        if not data_parts:
+            return None
+        data = "\n".join(data_parts).strip()
+        if not data:
+            return None
+        try:
+            result = json.loads(data)
+        except json.JSONDecodeError:
+            return None
+        return result if isinstance(result, dict) else None
+
     async def _request(
         self,
         method: str,

opencode_bridge/routes.py

@@ -49,7 +49,7 @@ def get_session_manager(handler: APIHandler) -> SessionManager:
 
 
 def _build_request_body(prompt: str, context: dict) -> dict:
-    """Build full request body for OpenCode POST /session/:id/message.
+    """Build full request body for OpenCode POST /session/:id/prompt_async.
 
     Returns dict with 'parts' (list) and 'system' (str) keys. No mode concept:
     the LLM interprets the user's natural-language instruction, with the code
@@ -105,6 +105,26 @@ def _strip_code_fence(s: str) -> str:
     return s
 
 
+def _extract_session_id(event: dict) -> str | None:
+    """Best-effort session ID extraction from an OpenCode event.
+
+    OpenCode's event shape is not strictly documented; this mirrors the
+    heuristics in demo.html ``handleGlobalEvent``:
+      properties.sessionID / properties.sessionId /
+      payload.properties.sessionID / syncEvent.aggregateID
+    """
+    props = event.get("properties") or {}
+    sid = props.get("sessionID") or props.get("sessionId")
+    if sid:
+        return sid
+    payload_props = (event.get("payload") or {}).get("properties") or {}
+    sid = payload_props.get("sessionID") or payload_props.get("sessionId")
+    if sid:
+        return sid
+    sync = event.get("syncEvent") or (event.get("payload") or {}).get("syncEvent") or {}
+    return sync.get("aggregateID")
+
+
 class HelloRouteHandler(APIHandler):
     # The following decorator should be present on all verb methods (head, get, post,
     # patch, put, delete, options) to ensure only authorized user can request the
@@ -154,6 +174,13 @@ class ProvidersHandler(APIHandler):
 
 
 class EditHandler(APIHandler):
+    """Async edit: fire prompt_async, return sessionId immediately.
+
+    The LLM's response is consumed via the separate ``/opencode-bridge/events``
+    SSE endpoint. Frontend subscribes to SSE after this call returns and
+    processes events incrementally.
+    """
+
     @tornado.web.authenticated
     async def post(self):
         try:
@@ -178,44 +205,174 @@ class EditHandler(APIHandler):
         try:
             sid = await sm.get_or_create(notebook_path)
             request_body = _build_request_body(prompt, context)
-            result = await client.send_message_sync(
+            await client.send_message_async(
                 sid,
                 request_body["parts"],
                 provider_id=provider_id,
                 model_id=model_id,
                 system=request_body["system"],
             )
-
-            text_parts = [
-                p.get("text", "")
-                for p in result.get("parts", [])
-                if p.get("type") == "text"
-            ]
-            # The AI's reply is markdown (code in ```fences```, optional
-            # explanation). Pass it through unchanged so the frontend
-            # `marked.parse` can render the code blocks and structure.
-            markdown = "\n".join(text_parts).strip()
-
             self.finish(json.dumps({
                 "ok": True,
-                "markdown": markdown,
                 "sessionId": sid,
                 "notebookPath": notebook_path,
             }))
         except OpenCodeError as e:
             # If session is invalid on OpenCode side, invalidate cache.
-            # 404 / 410 / "session not found" -> next get_or_create will recreate.
             if "404" in str(e) or "not found" in str(e).lower():
                 sm.invalidate(notebook_path)
                 log.warning("invalidated dead session for %s", notebook_path)
-            log.exception("edit failed")
+            log.exception("edit (async) failed")
             self.set_status(502)
             self.finish(json.dumps({"ok": False, "error": str(e)}))
         except Exception as e:
-            log.exception("edit failed")
+            log.exception("edit (async) failed")
+            self.set_status(502)
+            self.finish(json.dumps({"ok": False, "error": str(e)}))
+
+
+class GlobalEventHandler(APIHandler):
+    """SSE proxy for OpenCode Serve's ``GET /global/event``.
+
+    Streams every event from the upstream SSE feed to the client as
+    ``text/event-stream`` so EventSource / fetch+reader on the frontend
+    can consume them directly. The frontend filters by session ID —
+    we deliberately do NOT filter on the server because OpenCode's
+    event-shape has varied across versions and a too-eager server
+    filter can silently drop events the client would have accepted.
+
+    The connection is held open until the client disconnects, the upstream
+    closes, or an unrecoverable error occurs.
+    """
+
+    @tornado.web.authenticated
+    async def get(self):
+        # The ?session= query param is accepted for API compatibility
+        # but the server does NOT use it to filter — see docstring.
+        _ = self.get_query_argument("session", None)
+
+        self.set_header("Content-Type", "text/event-stream")
+        self.set_header("Cache-Control", "no-cache")
+        # Disable proxy buffering (e.g. nginx) so chunks reach the client
+        # immediately.
+        self.set_header("X-Accel-Buffering", "no")
+        # Prevent tornado from closing the connection on its own keep-alive
+        # timer — we hold it open as long as the upstream is alive.
+        self.set_header("Connection", "keep-alive")
+
+        client = make_client(self)
+        try:
+            async for event in client.stream_global_events():
+                payload = json.dumps(event, ensure_ascii=False)
+                self.write("data: %s\n\n" % payload)
+                # flush() pushes the chunk to the client; without it tornado
+                # would buffer until the response is "finished".
+                await self.flush()
+        except tornado.iostream.StreamClosedError:
+            # Client went away — normal disconnect, not an error.
+            log.info("SSE client disconnected")
+        except Exception as e:
+            log.exception("SSE proxy error")
+            try:
+                self.write("data: %s\n\n" % json.dumps({
+                    "type": "error",
+                    "message": str(e),
+                }))
+                await self.flush()
+            except Exception:
+                # Connection already gone; nothing to send.
+                pass
+
+
+class PermissionReplyHandler(APIHandler):
+    """Respond to a `permission.asked` event for a given session.
+
+    POST ``/opencode-bridge/permissions/:permId?session=:sid`` with body
+    ``{"response": "once"|"always"|"reject"}``. Forwards to OpenCode
+    Serve's ``POST /session/:sid/permissions/:permId``.
+    """
+
+    VALID_RESPONSES = ("once", "always", "reject")
+
+    @tornado.web.authenticated
+    async def post(self, perm_id: str):
+        session_id = self.get_query_argument("session", "")
+        if not session_id:
+            self.set_status(400)
+            self.finish(json.dumps({"error": "missing 'session' query parameter"}))
+            return
+        try:
+            body = json.loads(self.request.body) if self.request.body else {}
+        except json.JSONDecodeError as e:
+            self.set_status(400)
+            self.finish(json.dumps({"error": "bad request: %s" % e}))
+            return
+        response = body.get("response")
+        if response not in self.VALID_RESPONSES:
+            self.set_status(400)
+            self.finish(json.dumps({
+                "error": "response must be one of %s" % list(self.VALID_RESPONSES)
+            }))
+            return
+        client = make_client(self)
+        try:
+            ok = await client.reply_permission(session_id, perm_id, response)
+            self.finish(json.dumps({
+                "ok": ok,
+                "permissionId": perm_id,
+                "response": response,
+            }))
+        except OpenCodeError as e:
+            log.exception("permission reply failed")
+            self.set_status(502)
+            self.finish(json.dumps({"ok": False, "error": str(e)}))
+        except Exception as e:
+            log.exception("permission reply failed")
+            self.set_status(502)
+            self.finish(json.dumps({"ok": False, "error": str(e)}))
+
+
+class QuestionReplyHandler(APIHandler):
+    """Reply to a `question.asked` event with a freeform answer.
+
+    POST ``/opencode-bridge/questions/:qId/reply?session=:sid`` with body
+    ``{"answer": "..."}``. Forwards to OpenCode Serve's
+    ``POST /session/:sid/question/:qId/reply``.
+    """
+
+    @tornado.web.authenticated
+    async def post(self, question_id: str):
+        session_id = self.get_query_argument("session", "")
+        if not session_id:
+            self.set_status(400)
+            self.finish(json.dumps({"error": "missing 'session' query parameter"}))
+            return
+        try:
+            body = json.loads(self.request.body) if self.request.body else {}
+        except json.JSONDecodeError as e:
+            self.set_status(400)
+            self.finish(json.dumps({"error": "bad request: %s" % e}))
+            return
+        answer = body.get("answer")
+        if not isinstance(answer, str) or not answer.strip():
+            self.set_status(400)
+            self.finish(json.dumps({"error": "missing 'answer'"}))
+            return
+        client = make_client(self)
+        try:
+            ok = await client.reply_question(session_id, question_id, answer)
+            self.finish(json.dumps({
+                "ok": ok,
+                "questionId": question_id,
+            }))
+        except OpenCodeError as e:
+            log.exception("question reply failed")
+            self.set_status(502)
+            self.finish(json.dumps({"ok": False, "error": str(e)}))
+        except Exception as e:
+            log.exception("question reply failed")
             self.set_status(502)
             self.finish(json.dumps({"ok": False, "error": str(e)}))
-        # NO finally delete — session is reused per notebook.
 
 
 class SessionListHandler(APIHandler):
@@ -310,9 +467,21 @@ def setup_route_handlers(web_app):
         (url_path_join(base_url, "opencode-bridge", "health"), HealthHandler),
         (url_path_join(base_url, "opencode-bridge", "providers"), ProvidersHandler),
         (url_path_join(base_url, "opencode-bridge", "edit"), EditHandler),
+        (url_path_join(base_url, "opencode-bridge", "events"), GlobalEventHandler),
         (url_path_join(base_url, "opencode-bridge", "sessions"), SessionListHandler),
         (url_path_join(base_url, "opencode-bridge", "session"), SessionReleaseHandler),
         (url_path_join(base_url, "opencode-bridge", "session-messages"), SessionMessagesHandler),
+        # /permissions/<perm_id>  and  /questions/<q_id>/reply
+        # use a URLSpec with a regex so the path param is forwarded to
+        # the handler as a positional argument.
+        tornado.web.URLSpec(
+            url_path_join(base_url, "opencode-bridge", "permissions", r"([A-Za-z0-9_\-]+)"),
+            PermissionReplyHandler,
+        ),
+        tornado.web.URLSpec(
+            url_path_join(base_url, "opencode-bridge", "questions", r"([A-Za-z0-9_\-]+)", "reply"),
+            QuestionReplyHandler,
+        ),
     ]
 
     web_app.add_handlers(host_pattern, handlers)

opencode_bridge/tests/test_client.py

@@ -98,28 +98,30 @@ async def test_auth_absent_when_no_password(base_config: OpenCodeConfig) -> None
 
 
 @pytest.mark.asyncio
-async def test_send_message_sync_includes_model_when_provider_and_model_given(
+async def test_send_message_async_includes_model_when_provider_and_model_given(
     base_config: OpenCodeConfig,
 ) -> None:
+    """send_message_async posts to /prompt_async with provider/model when both set."""
     client, mock = _make_client(base_config, [(200, {"done": True})])
     parts = [{"type": "text", "text": "hello"}]
-    result = await client.send_message_sync("s1", parts, provider_id="p1", model_id="m1")
-    assert result == {"done": True}
+    result = await client.send_message_async("s1", parts, provider_id="p1", model_id="m1")
+    assert result is None
     call = mock.calls[0]
-    assert call["request"].url == "http://127.0.0.1:4096/session/s1/message"
+    assert call["request"].url == "http://127.0.0.1:4096/session/s1/prompt_async"
+    assert call["request"].method == "POST"
     body = json.loads(call["request"].body.decode("utf-8"))
     assert body["parts"] == parts
     assert body["model"] == {"providerID": "p1", "modelID": "m1"}
 
 
 @pytest.mark.asyncio
-async def test_send_message_sync_omits_model_when_provider_or_model_missing(
+async def test_send_message_async_omits_model_when_provider_or_model_missing(
     base_config: OpenCodeConfig,
 ) -> None:
     client, mock = _make_client(base_config, [(200, {"done": True})])
     parts = [{"type": "text", "text": "hello"}]
-    result = await client.send_message_sync("s1", parts)
-    assert result == {"done": True}
+    result = await client.send_message_async("s1", parts)
+    assert result is None
     call = mock.calls[0]
     body = json.loads(call["request"].body.decode("utf-8"))
     assert body["parts"] == parts
@@ -127,7 +129,7 @@ async def test_send_message_sync_omits_model_when_provider_or_model_missing(
 
 
 @pytest.mark.asyncio
-async def test_send_message_sync_includes_system_when_provided() -> None:
+async def test_send_message_async_includes_system_when_provided() -> None:
     """system param is forwarded into the request body when not None."""
     config = OpenCodeConfig(
         url="http://x:1", user="u", password="", request_timeout_seconds=120
@@ -136,7 +138,7 @@ async def test_send_message_sync_includes_system_when_provided() -> None:
     mock.responses.append((200, {"info": {}, "parts": []}))
     client = OpenCodeClient(config, http_client=mock)
 
-    await client.send_message_sync("sid", [{"type": "text", "text": "hi"}], system="be brief")
+    await client.send_message_async("sid", [{"type": "text", "text": "hi"}], system="be brief")
 
     assert len(mock.calls) == 1
     body = json.loads(mock.calls[0]["request"].body.decode("utf-8"))
@@ -144,6 +146,83 @@ async def test_send_message_sync_includes_system_when_provided() -> None:
     assert body["parts"] == [{"type": "text", "text": "hi"}]
 
 
+@pytest.mark.asyncio
+async def test_stream_global_events_yields_parsed_events(
+    base_config: OpenCodeConfig,
+) -> None:
+    """stream_global_events yields one dict per `data: {...}` SSE block."""
+
+    class _StreamingMock(MockHTTPClient):
+        async def fetch(self, request, **kwargs):
+            self.calls.append({"request": request, "kwargs": kwargs})
+            # Simulate OpenCode's /global/event: three events back-to-back.
+            chunks = [
+                b"data: {\"type\": \"session.next.text.delta\", \"properties\": "
+                b"{\"sessionID\": \"s1\", \"delta\": \"hello \"}}\n\n",
+                b"data: {\"type\": \"session.next.text.delta\", \"properties\": "
+                b"{\"sessionID\": \"s1\", \"delta\": \"world\"}}\n\n",
+                b"data: {\"type\": \"session.idle\", \"properties\": "
+                b"{\"sessionID\": \"s1\"}}\n\n",
+            ]
+            # streaming_callback now lives on the HTTPRequest (per tornado
+            # API — fetch() rejects kwargs when an HTTPRequest is passed).
+            cb = getattr(request, "streaming_callback", None)
+            if cb is not None:
+                for c in chunks:
+                    cb(c)
+            # Empty body, code 200 to terminate normally.
+            return tornado.httpclient.HTTPResponse(
+                request=request,
+                code=200,
+                headers=tornado.httputil.HTTPHeaders(),
+                buffer=BytesIO(b""),
+            )
+
+    mock = _StreamingMock()
+    client = OpenCodeClient(base_config, http_client=mock)
+
+    events = []
+    async for ev in client.stream_global_events():
+        events.append(ev)
+
+    assert len(events) == 3
+    assert events[0]["type"] == "session.next.text.delta"
+    assert events[0]["properties"]["delta"] == "hello "
+    assert events[1]["properties"]["delta"] == "world"
+    assert events[2]["type"] == "session.idle"
+    # Stream used the SSE accept header and no timeout.
+    call = mock.calls[0]
+    assert call["kwargs"] == {"raise_error": False}
+    assert call["request"].method == "GET"
+    assert call["request"].url == "http://127.0.0.1:4096/global/event"
+    assert call["request"].request_timeout == 0
+    assert call["request"].headers.get("Accept") == "text/event-stream"
+    assert call["request"].streaming_callback is not None
+
+
+def test_parse_sse_block_handles_concatenated_data_lines() -> None:
+    """Multi-line `data:` blocks are joined with \n (per SSE spec)."""
+    parsed = OpenCodeClient._parse_sse_block(
+        "data: {\"a\": 1,\ndata:  \"b\": 2}"
+    )
+    assert parsed == {"a": 1, "b": 2}
+
+
+def test_parse_sse_block_ignores_non_data_fields() -> None:
+    parsed = OpenCodeClient._parse_sse_block("event: ping\ndata: {\"k\": \"v\"}\n\n")
+    assert parsed == {"k": "v"}
+
+
+def test_parse_sse_block_returns_none_on_empty() -> None:
+    assert OpenCodeClient._parse_sse_block("") is None
+    assert OpenCodeClient._parse_sse_block("event: ping\n\n") is None
+    assert OpenCodeClient._parse_sse_block(": heartbeat\n\n") is None
+
+
+def test_parse_sse_block_returns_none_on_bad_json() -> None:
+    assert OpenCodeClient._parse_sse_block("data: not-json\n\n") is None
+
+
 @pytest.mark.asyncio
 async def test_delete_session_returns_bool_by_status(base_config: OpenCodeConfig) -> None:
     client, mock = _make_client(base_config, [(200, True), (404, False)])
@@ -154,6 +233,39 @@ async def test_delete_session_returns_bool_by_status(base_config: OpenCodeConfig
     assert mock.calls[0]["request"].url == "http://127.0.0.1:4096/session/s1"
 
 
+@pytest.mark.asyncio
+async def test_reply_permission_posts_response(base_config: OpenCodeConfig) -> None:
+    """reply_permission forwards the response string to /session/:id/permissions/:permId."""
+    client, mock = _make_client(base_config, [(200, True)])
+    ok = await client.reply_permission("s1", "perm-1", "once")
+    assert ok is True
+    call = mock.calls[0]
+    assert call["request"].url == "http://127.0.0.1:4096/session/s1/permissions/perm-1"
+    assert call["request"].method == "POST"
+    body = json.loads(call["request"].body.decode("utf-8"))
+    assert body == {"response": "once"}
+
+
+@pytest.mark.asyncio
+async def test_reply_permission_returns_false_on_404(
+    base_config: OpenCodeConfig,
+) -> None:
+    client, mock = _make_client(base_config, [(404, False)])
+    assert await client.reply_permission("s1", "perm-1", "reject") is False
+
+
+@pytest.mark.asyncio
+async def test_reply_question_posts_answer(base_config: OpenCodeConfig) -> None:
+    """reply_question forwards the answer to /session/:id/question/:qid/reply."""
+    client, mock = _make_client(base_config, [(200, True)])
+    ok = await client.reply_question("s1", "q-7", "use 3.12")
+    assert ok is True
+    call = mock.calls[0]
+    assert call["request"].url == "http://127.0.0.1:4096/session/s1/question/q-7/reply"
+    body = json.loads(call["request"].body.decode("utf-8"))
+    assert body == {"answer": "use 3.12"}
+
+
 @pytest.mark.asyncio
 async def test_non_2xx_response_raises_with_body(base_config: OpenCodeConfig) -> None:
     client, mock = _make_client(base_config, [(500, {"error": "boom"})])

opencode_bridge/tests/test_routes.py

@@ -12,10 +12,6 @@ class FakeOpenCodeClient:
         self.session_id = "fake-session-123"
         self.health_response = {"healthy": True, "version": "0.1.0"}
         self.providers_response = {"providers": [{"id": "anthropic", "models": [{"id": "claude"}]}]}
-        self.message_response = {
-            "info": {"id": "msg-1"},
-            "parts": [{"type": "text", "text": "def foo():\n    return 42\n"}],
-        }
         # Canned session-messages list response (list_session_messages).
         # Default: one user + one assistant message, mixed text parts.
         self.messages_response = [
@@ -45,9 +41,13 @@ class FakeOpenCodeClient:
         self.calls.append(("create_session", title))
         return {"id": self.session_id, "title": title}
 
-    async def send_message_sync(self, session_id, parts, provider_id=None, model_id=None, system=None):
-        self.calls.append(("send_message_sync", session_id, parts, system))
-        return self.message_response
+    async def send_message_async(
+        self, session_id, parts, provider_id=None, model_id=None, system=None
+    ):
+        self.calls.append(
+            ("send_message_async", session_id, parts, system, provider_id, model_id)
+        )
+        return None
 
     async def delete_session(self, session_id):
         self.calls.append(("delete_session", session_id))
@@ -57,6 +57,19 @@ class FakeOpenCodeClient:
         self.calls.append(("list_session_messages", session_id))
         return self.messages_response
 
+    async def reply_permission(self, session_id, permission_id, response):
+        self.calls.append(("reply_permission", session_id, permission_id, response))
+        return True
+
+    async def reply_question(self, session_id, question_id, answer):
+        self.calls.append(("reply_question", session_id, question_id, answer))
+        return True
+
+    async def stream_global_events(self):
+        """Async generator — tests can monkey-patch this to control the stream."""
+        if False:  # pragma: no cover - never executed, just makes this an asyncgen
+            yield {}
+
 
 class FakeSessionManager:
     """Drop-in replacement for SessionManager with recording."""
@@ -131,6 +144,12 @@ async def test_providers_handler(monkeypatch, jp_fetch):
 
 
 async def test_edit_handler(monkeypatch, jp_fetch):
+    """POST /opencode-bridge/edit fires prompt_async and returns sessionId only.
+
+    The async path returns immediately with `{ok, sessionId, notebookPath}` —
+    the actual LLM response is consumed via the /events SSE endpoint, not
+    embedded in this response.
+    """
     fake = FakeOpenCodeClient()
     monkeypatch.setattr("opencode_bridge.routes.make_client", lambda h: fake)
 
@@ -159,31 +178,68 @@ async def test_edit_handler(monkeypatch, jp_fetch):
     )
     assert response.code == 200
     payload = json.loads(response.body)
-    assert payload["ok"] is True
-    # Server now returns the AI reply as raw markdown (```fences``` kept
-    # so the frontend marked.parse renders code blocks).
-    assert payload["markdown"] == "def foo():\n    return 42"
-    assert payload["sessionId"] == "fake-session-123"
-    assert payload["notebookPath"] == "test.ipynb"
+    assert payload == {
+        "ok": True,
+        "sessionId": "fake-session-123",
+        "notebookPath": "test.ipynb",
+    }
+    # Async path: no `markdown` field. The LLM reply is consumed via SSE.
+    assert "markdown" not in payload
 
-    # Session manager was used, not direct create/delete on client
+    # Session manager was used (not direct create/delete on client).
     call_names = [c[0] for c in fake.calls]
     assert "create_session" not in call_names
     assert "delete_session" not in call_names
-    assert "send_message_sync" in call_names
+    assert "send_message_async" in call_names
+    assert "send_message_sync" not in call_names  # legacy sync path is gone
 
-    # send_message_sync received the session ID from the manager
-    send_call = [c for c in fake.calls if c[0] == "send_message_sync"][0]
+    # send_message_async received the session ID from the manager.
+    send_call = [c for c in fake.calls if c[0] == "send_message_async"][0]
     assert send_call[1] == "fake-session-123"
-    # system prompt was passed (v3+ allows markdown/fences in the reply)
+    # system prompt was passed.
     assert "你是一个代码助手" in send_call[3]
 
-    # SessionManager.get_or_create was called with the notebook path
+    # SessionManager.get_or_create was called with the notebook path.
     sm_call_names = [c[0] for c in fake_sm.calls]
     assert sm_call_names == ["get_or_create"]
     assert fake_sm.calls[0][1] == "test.ipynb"
 
 
+async def test_edit_handler_includes_model_when_provider_and_model_given(
+    monkeypatch, jp_fetch
+):
+    """providerId/modelId are forwarded into the async send body when both set."""
+    fake = FakeOpenCodeClient()
+    monkeypatch.setattr("opencode_bridge.routes.make_client", lambda h: fake)
+    monkeypatch.setattr(
+        "opencode_bridge.routes.get_session_manager",
+        lambda h: FakeSessionManager(),
+    )
+
+    body = json.dumps({
+        "prompt": "x",
+        "context": {
+            "notebookPath": "n.ipynb",
+            "cellId": "c",
+            "language": "python",
+            "cellIndex": 0,
+            "totalCells": 1,
+            "source": "",
+            "previousCode": None,
+            "error": None,
+        },
+        "providerId": "anthropic",
+        "modelId": "claude-sonnet-4-20250514",
+    })
+    response = await jp_fetch("opencode-bridge", "edit", method="POST", body=body)
+    assert response.code == 200
+
+    send_call = [c for c in fake.calls if c[0] == "send_message_async"][0]
+    # (send_message_async, sid, parts, system, provider_id, model_id)
+    assert send_call[4] == "anthropic"
+    assert send_call[5] == "claude-sonnet-4-20250514"
+
+
 async def test_session_list_handler(monkeypatch, jp_fetch):
     fake_sm = FakeSessionManager()
     fake_sm._sessions = {
@@ -287,3 +343,119 @@ async def test_session_release_handler(monkeypatch, jp_fetch):
     assert payload["notebookPath"] == "foo.ipynb"
     assert payload["deleted"] is True
     assert ("release", "foo.ipynb") in fake_sm.calls
+
+
+async def test_permission_reply_forwards_response(monkeypatch, jp_fetch):
+    """POST /opencode-bridge/permissions/:permId?session=:sid forwards to client."""
+    fake = FakeOpenCodeClient()
+    monkeypatch.setattr("opencode_bridge.routes.make_client", lambda h: fake)
+
+    response = await jp_fetch(
+        "opencode-bridge", "permissions", "perm-42",
+        method="POST",
+        params={"session": "ses-1"},
+        body=json.dumps({"response": "once"}),
+    )
+    assert response.code == 200
+    payload = json.loads(response.body)
+    assert payload == {
+        "ok": True,
+        "permissionId": "perm-42",
+        "response": "once",
+    }
+    assert ("reply_permission", "ses-1", "perm-42", "once") in fake.calls
+
+
+async def test_permission_reply_rejects_invalid_response(monkeypatch, jp_fetch):
+    """response must be one of once/always/reject — 400 otherwise."""
+    fake = FakeOpenCodeClient()
+    monkeypatch.setattr("opencode_bridge.routes.make_client", lambda h: fake)
+
+    with pytest.raises(tornado.httpclient.HTTPClientError) as exc_info:
+        await jp_fetch(
+            "opencode-bridge", "permissions", "perm-1",
+            method="POST",
+            params={"session": "ses-1"},
+            body=json.dumps({"response": "yes"}),
+        )
+    assert exc_info.value.code == 400
+    assert "yes" not in fake.calls and fake.calls == []
+
+
+async def test_permission_reply_requires_session(monkeypatch, jp_fetch):
+    """Missing ?session=... is a 400."""
+    fake = FakeOpenCodeClient()
+    monkeypatch.setattr("opencode_bridge.routes.make_client", lambda h: fake)
+    with pytest.raises(tornado.httpclient.HTTPClientError) as exc_info:
+        await jp_fetch(
+            "opencode-bridge", "permissions", "perm-1",
+            method="POST",
+            body=json.dumps({"response": "once"}),
+        )
+    assert exc_info.value.code == 400
+
+
+async def test_question_reply_forwards_answer(monkeypatch, jp_fetch):
+    fake = FakeOpenCodeClient()
+    monkeypatch.setattr("opencode_bridge.routes.make_client", lambda h: fake)
+
+    response = await jp_fetch(
+        "opencode-bridge", "questions", "q-7", "reply",
+        method="POST",
+        params={"session": "ses-2"},
+        body=json.dumps({"answer": "use Python 3.12"}),
+    )
+    assert response.code == 200
+    payload = json.loads(response.body)
+    assert payload == {"ok": True, "questionId": "q-7"}
+    assert ("reply_question", "ses-2", "q-7", "use Python 3.12") in fake.calls
+
+
+async def test_question_reply_rejects_empty_answer(monkeypatch, jp_fetch):
+    fake = FakeOpenCodeClient()
+    monkeypatch.setattr("opencode_bridge.routes.make_client", lambda h: fake)
+    with pytest.raises(tornado.httpclient.HTTPClientError) as exc_info:
+        await jp_fetch(
+            "opencode-bridge", "questions", "q-1", "reply",
+            method="POST",
+            params={"session": "ses-1"},
+            body=json.dumps({"answer": "   "}),
+        )
+    assert exc_info.value.code == 400
+    assert fake.calls == []
+
+
+async def test_global_event_handler_forwards_all_events_unfiltered(
+    monkeypatch, jp_fetch
+):
+    """Server-side SSE proxy must NOT filter by session — it forwards
+    everything so the client can decide. (Previously a too-eager server
+    filter silently dropped events the client would have accepted.)
+
+    We mock stream_global_events to push three events with different
+    sessionID shapes (one matching, one not, one missing). The proxy
+    should pass all three through as data: lines.
+    """
+    fake = FakeOpenCodeClient()
+
+    async def fake_stream():
+        yield {"type": "session.next.text.delta", "properties": {"sessionID": "ses-1", "delta": "A"}}
+        yield {"type": "session.next.text.delta", "properties": {"sessionID": "ses-OTHER", "delta": "B"}}
+        yield {"type": "server.connected", "properties": {}}
+
+    fake.stream_global_events = fake_stream
+    monkeypatch.setattr("opencode_bridge.routes.make_client", lambda h: fake)
+
+    # Use ?session=ses-1 to confirm the param is accepted but doesn't filter.
+    response = await jp_fetch(
+        "opencode-bridge", "events",
+        params={"session": "ses-1"},
+    )
+    assert response.code == 200
+    body = response.body.decode("utf-8")
+    # All three events should appear in the body, regardless of sessionID.
+    assert '"delta": "A"' in body
+    assert '"delta": "B"' in body
+    assert '"server.connected"' in body
+    # The body should use the SSE format: each event as a `data:` line.
+    assert body.count("data: ") == 3

src/__tests__/model_selection.spec.ts

@@ -0,0 +1,99 @@
+/**
+ * Tests for the localStorage-backed model selection persistence helper.
+ * jsdom provides a real (per-test-class) localStorage.
+ */
+import {
+  clearModelSelection,
+  loadModelSelection,
+  saveModelSelection
+} from '../components/model_selection';
+
+describe('loadModelSelection', () => {
+  beforeEach(() => {
+    localStorage.clear();
+  });
+
+  it('returns null when no entry exists', () => {
+    expect(loadModelSelection('foo.ipynb')).toBeNull();
+  });
+
+  it('returns null for an empty notebookPath (defensive)', () => {
+    expect(loadModelSelection('')).toBeNull();
+  });
+
+  it('returns the parsed entry when storage is valid', () => {
+    localStorage.setItem(
+      'opencode_bridge:model-selection:foo.ipynb',
+      JSON.stringify({ providerId: 'anthropic', modelId: 'claude-sonnet-4-20250514' })
+    );
+    expect(loadModelSelection('foo.ipynb')).toEqual({
+      providerId: 'anthropic',
+      modelId: 'claude-sonnet-4-20250514'
+    });
+  });
+
+  it('returns null for malformed JSON (does not throw)', () => {
+    localStorage.setItem(
+      'opencode_bridge:model-selection:foo.ipynb',
+      '{not json'
+    );
+    expect(loadModelSelection('foo.ipynb')).toBeNull();
+  });
+
+  it('returns null when the entry is missing required fields', () => {
+    localStorage.setItem(
+      'opencode_bridge:model-selection:foo.ipynb',
+      JSON.stringify({ providerId: 'anthropic' })
+    );
+    expect(loadModelSelection('foo.ipynb')).toBeNull();
+  });
+
+  it('keys entries independently per notebookPath', () => {
+    saveModelSelection('a.ipynb', { providerId: 'p1', modelId: 'm1' });
+    saveModelSelection('b.ipynb', { providerId: 'p2', modelId: 'm2' });
+    expect(loadModelSelection('a.ipynb')).toEqual({
+      providerId: 'p1',
+      modelId: 'm1'
+    });
+    expect(loadModelSelection('b.ipynb')).toEqual({
+      providerId: 'p2',
+      modelId: 'm2'
+    });
+  });
+});
+
+describe('saveModelSelection + clearModelSelection', () => {
+  beforeEach(() => {
+    localStorage.clear();
+  });
+
+  it('writes a JSON entry that can be re-read', () => {
+    saveModelSelection('foo.ipynb', { providerId: 'p1', modelId: 'm1' });
+    expect(loadModelSelection('foo.ipynb')).toEqual({
+      providerId: 'p1',
+      modelId: 'm1'
+    });
+  });
+
+  it('overwrites an existing entry', () => {
+    saveModelSelection('foo.ipynb', { providerId: 'p1', modelId: 'm1' });
+    saveModelSelection('foo.ipynb', { providerId: 'p2', modelId: 'm2' });
+    expect(loadModelSelection('foo.ipynb')).toEqual({
+      providerId: 'p2',
+      modelId: 'm2'
+    });
+  });
+
+  it('clearModelSelection removes the entry', () => {
+    saveModelSelection('foo.ipynb', { providerId: 'p1', modelId: 'm1' });
+    clearModelSelection('foo.ipynb');
+    expect(loadModelSelection('foo.ipynb')).toBeNull();
+  });
+
+  it('no-op for an empty notebookPath (defensive)', () => {
+    expect(() =>
+      saveModelSelection('', { providerId: 'p1', modelId: 'm1' })
+    ).not.toThrow();
+    expect(() => clearModelSelection('')).not.toThrow();
+  });
+});

src/__tests__/opencode_cell_actions.spec.ts

@@ -103,13 +103,22 @@ beforeAll(() => {
   });
 });
 
+beforeEach(() => {
+  // Each test starts with a clean selection store so persistence tests
+  // are deterministic.
+  localStorage.clear();
+});
+
 // Mock the network-touching API module so jsdom tests don't try to hit
 // a real notebook server (and to keep the refresh-history logic decoupled
 // from the network). The cell_actions tests focus on widget behavior.
 jest.mock('../api/opencode_client', () => ({
   callOpenCodeEdit: jest.fn(),
   callOpenCodeProviders: jest.fn(),
-  callOpenCodeSessionMessages: jest.fn().mockResolvedValue({ messages: [] })
+  callOpenCodeSessionMessages: jest.fn().mockResolvedValue({ messages: [] }),
+  callOpenCodeReplyPermission: jest.fn().mockResolvedValue({ ok: true }),
+  callOpenCodeReplyQuestion: jest.fn().mockResolvedValue({ ok: true }),
+  subscribeOpenCodeEvents: jest.fn().mockReturnValue({ close: jest.fn() })
 }));
 
 import {
@@ -119,6 +128,20 @@ import {
 } from '../components/opencode_cell_actions';
 import { OpenCodeInlinePrompt } from '../components/opencode_inline_prompt';
 import { CodeCell } from '@jupyterlab/cells';
+import {
+  callOpenCodeEdit,
+  subscribeOpenCodeEvents
+} from '../api/opencode_client';
+import { Notification } from '@jupyterlab/apputils';
+
+const mockedCallOpenCodeEdit = callOpenCodeEdit as jest.MockedFunction<
+  typeof callOpenCodeEdit
+>;
+const mockedSubscribeOpenCodeEvents =
+  subscribeOpenCodeEvents as jest.MockedFunction<
+    typeof subscribeOpenCodeEvents
+  >;
+const mockedNotification = Notification as jest.Mocked<typeof Notification>;
 
 function makeFakeOutputs(items: any[]): any {
   return {
@@ -306,6 +329,122 @@ describe('OpenCodeCellActions', () => {
     setOpenCodeProviders(null);
   });
 
+  it('restores the user\'s previously-saved (provider, model) when still valid in the providers payload', () => {
+    // Seed localStorage with a prior selection that IS valid.
+    localStorage.setItem(
+      'opencode_bridge:model-selection:foo.ipynb',
+      JSON.stringify({ providerId: 'openai', modelId: 'gpt-5' })
+    );
+    setOpenCodeProviders({
+      providers: [
+        { id: 'anthropic', models: { 'claude-sonnet-4-20250514': { id: 'x' } } },
+        { id: 'openai', models: { 'gpt-5': { id: 'x' }, 'gpt-4': { id: 'x' } } }
+      ],
+      default: { anthropic: 'claude-sonnet-4-20250514' }
+    });
+    const cell = makeFakeCell('x = 1', [], 'foo.ipynb');
+    const actions = new OpenCodeCellActions(cell);
+    Object.defineProperty(actions, 'parent', { value: cell, configurable: true });
+    (actions as any).onAfterAttach({} as any);
+
+    (actions.node.querySelector('button') as HTMLButtonElement).click();
+    const prompt = (actions as any)._prompt;
+    const providerSel = prompt.node.querySelector(
+      '.opencode-provider-select'
+    ) as HTMLSelectElement;
+    const modelSel = prompt.node.querySelector(
+      '.opencode-model-select'
+    ) as HTMLSelectElement;
+    // Stored selection was applied, NOT the default (first provider).
+    expect(providerSel.value).toBe('openai');
+    expect(modelSel.value).toBe('gpt-5');
+    setOpenCodeProviders(null);
+  });
+
+  it('falls back to default when the stored provider is no longer present', () => {
+    localStorage.setItem(
+      'opencode_bridge:model-selection:foo.ipynb',
+      JSON.stringify({ providerId: 'removed-provider', modelId: 'm1' })
+    );
+    setOpenCodeProviders({
+      providers: [
+        { id: 'anthropic', models: { 'claude-sonnet-4-20250514': { id: 'x' } } }
+      ]
+    });
+    const cell = makeFakeCell('x = 1', [], 'foo.ipynb');
+    const actions = new OpenCodeCellActions(cell);
+    Object.defineProperty(actions, 'parent', { value: cell, configurable: true });
+    (actions as any).onAfterAttach({} as any);
+
+    (actions.node.querySelector('button') as HTMLButtonElement).click();
+    const providerSel = (actions as any)._prompt.node.querySelector(
+      '.opencode-provider-select'
+    ) as HTMLSelectElement;
+    // Provider is gone → fall back to first available.
+    expect(providerSel.value).toBe('anthropic');
+    setOpenCodeProviders(null);
+  });
+
+  it('falls back to default model when the stored model is no longer present under the same provider', () => {
+    localStorage.setItem(
+      'opencode_bridge:model-selection:foo.ipynb',
+      JSON.stringify({ providerId: 'openai', modelId: 'gpt-99-removed' })
+    );
+    setOpenCodeProviders({
+      providers: [
+        { id: 'openai', models: { 'gpt-5': { id: 'x' }, 'gpt-4': { id: 'x' } } }
+      ],
+      default: { openai: 'gpt-5' }
+    });
+    const cell = makeFakeCell('x = 1', [], 'foo.ipynb');
+    const actions = new OpenCodeCellActions(cell);
+    Object.defineProperty(actions, 'parent', { value: cell, configurable: true });
+    (actions as any).onAfterAttach({} as any);
+
+    (actions.node.querySelector('button') as HTMLButtonElement).click();
+    const modelSel = (actions as any)._prompt.node.querySelector(
+      '.opencode-model-select'
+    ) as HTMLSelectElement;
+    // Model is gone → fall back to default[openai] = gpt-5.
+    expect(modelSel.value).toBe('gpt-5');
+    setOpenCodeProviders(null);
+  });
+
+  it('saves the new selection to localStorage when the model select changes', () => {
+    setOpenCodeProviders({
+      providers: [
+        { id: 'anthropic', models: { 'claude-sonnet-4-20250514': { id: 'x' } } },
+        { id: 'openai', models: { 'gpt-5': { id: 'x' } } }
+      ]
+    });
+    const cell = makeFakeCell('x = 1', [], 'foo.ipynb');
+    const actions = new OpenCodeCellActions(cell);
+    Object.defineProperty(actions, 'parent', { value: cell, configurable: true });
+    (actions as any).onAfterAttach({} as any);
+
+    (actions.node.querySelector('button') as HTMLButtonElement).click();
+    const prompt = (actions as any)._prompt;
+    const providerSel = prompt.node.querySelector(
+      '.opencode-provider-select'
+    ) as HTMLSelectElement;
+    const modelSel = prompt.node.querySelector(
+      '.opencode-model-select'
+    ) as HTMLSelectElement;
+
+    // Switch provider to openai (whose model list includes gpt-5).
+    providerSel.value = 'openai';
+    providerSel.dispatchEvent(new Event('change'));
+    // Now change the model select to gpt-5.
+    modelSel.value = 'gpt-5';
+    modelSel.dispatchEvent(new Event('change'));
+
+    const stored = JSON.parse(
+      localStorage.getItem('opencode_bridge:model-selection:foo.ipynb') || 'null'
+    );
+    expect(stored).toEqual({ providerId: 'openai', modelId: 'gpt-5' });
+    setOpenCodeProviders(null);
+  });
+
   it('falls back to the first model when default[provider] is not in models', () => {
     setOpenCodeProviders({
       providers: [
@@ -395,58 +534,197 @@ describe('OpenCodeCellActions', () => {
     ).toBeNull();
   });
 
-  it('on a successful response, does NOT replace the cell source and triggers a history refresh', () => {
+  it('on submit: posts to /edit, binds SSE, does NOT replace cell source, clears input', async () => {
     setOpenCodeProviders(null);
-    const cell = makeFakeCell('x = 1');
-    const actions = new OpenCodeCellActions(cell);
-    // Click once to create the inline prompt (so the history refresh
-    // has a target and the cell source check has a baseline).
-    Object.defineProperty(actions, 'parent', { value: cell, configurable: true });
-    (actions as any).onAfterAttach({} as any);
-    const btn = actions.node.querySelector('button') as HTMLButtonElement;
-    btn.click();
-
-    // The history refresh on show is async; spy on _refreshHistory to
-    // confirm it gets called (without depending on the network).
-    const refreshSpy = jest.spyOn(actions as any, '_refreshHistory');
-
-    (actions as any)._handleResponse({
+    mockedCallOpenCodeEdit.mockResolvedValue({
       ok: true,
-      markdown: '# AI says hi\n\n```python\nprint("hi")\n```',
       sessionId: 'ses-abc',
       notebookPath: 'foo.ipynb'
     });
+    const subscribeCalls: any[] = [];
+    mockedSubscribeOpenCodeEvents.mockImplementation(
+      ((sid: string, _s: any, handlers: any) => {
+        subscribeCalls.push({ sid, handlers });
+        return { close: jest.fn() };
+      }) as any
+    );
 
-    // The cell source must remain unchanged.
-    expect((cell as any).model.sharedModel.setSource).not.toHaveBeenCalled();
-    // A history refresh was triggered (to pick up the new assistant msg).
-    expect(refreshSpy).toHaveBeenCalledWith('foo.ipynb');
-  });
-
-  it('clears the input textarea after a successful response', () => {
-    setOpenCodeProviders(null);
     const cell = makeFakeCell('x = 1');
     const actions = new OpenCodeCellActions(cell);
     Object.defineProperty(actions, 'parent', { value: cell, configurable: true });
     (actions as any).onAfterAttach({} as any);
-
-    // Open the prompt and type something into the textarea.
     const aiBtn = actions.node.querySelector('button') as HTMLButtonElement;
     aiBtn.click();
+
+    // Type into the textarea, then submit.
     const prompt = (actions as any)._prompt;
-    expect(prompt).not.toBeNull();
     const textarea = prompt._textarea as HTMLTextAreaElement;
     textarea.value = 'make it faster';
+    (prompt._sendBtn as HTMLButtonElement).click();
 
-    // A successful response clears the input so the user can type a
-    // follow-up message without manually deleting the previous prompt.
-    (actions as any)._handleResponse({
+    // Let the async _onSubmit run.
+    await new Promise(r => setTimeout(r, 5));
+
+    // callOpenCodeEdit was called with the right body.
+    expect(mockedCallOpenCodeEdit).toHaveBeenCalledTimes(1);
+    expect(mockedCallOpenCodeEdit.mock.calls[0][0].prompt).toBe('make it faster');
+    expect(mockedCallOpenCodeEdit.mock.calls[0][0].context.notebookPath).toBe('foo.ipynb');
+
+    // SSE subscription was opened with the returned sessionId.
+    expect(subscribeCalls).toHaveLength(1);
+    expect(subscribeCalls[0].sid).toBe('ses-abc');
+    expect(typeof subscribeCalls[0].handlers.onEvent).toBe('function');
+
+    // The cell source must remain unchanged.
+    expect((cell as any).model.sharedModel.setSource).not.toHaveBeenCalled();
+
+    // The textarea is cleared so the user can type a follow-up.
+    expect(textarea.value).toBe('');
+  });
+
+  it('on submit: forwards SSE events to the prompt and closes the stream on session.idle', async () => {
+    setOpenCodeProviders(null);
+    mockedCallOpenCodeEdit.mockResolvedValue({
       ok: true,
-      markdown: 'ok',
       sessionId: 'ses-1',
       notebookPath: 'foo.ipynb'
     });
-    expect(textarea.value).toBe('');
+
+    const fakeSse = { close: jest.fn() };
+    let captured: any = null;
+    mockedSubscribeOpenCodeEvents.mockImplementation(
+      ((_sid: string, _s: any, handlers: any) => {
+        captured = handlers;
+        return fakeSse;
+      }) as any
+    );
+
+    const cell = makeFakeCell('x = 1');
+    const actions = new OpenCodeCellActions(cell);
+    Object.defineProperty(actions, 'parent', { value: cell, configurable: true });
+    (actions as any).onAfterAttach({} as any);
+    const aiBtn = actions.node.querySelector('button') as HTMLButtonElement;
+    aiBtn.click();
+    const prompt = (actions as any)._prompt;
+    const textarea = prompt._textarea as HTMLTextAreaElement;
+    textarea.value = 'go';
+    (prompt._sendBtn as HTMLButtonElement).click();
+
+    await new Promise(r => setTimeout(r, 5));
+    expect(captured).not.toBeNull();
+
+    // Simulate a text-delta event arriving.
+    const applySpy = jest.spyOn(prompt, 'applyEvent');
+    captured.onEvent({
+      type: 'session.next.text.delta',
+      properties: { sessionID: 'ses-1', delta: 'hel' }
+    });
+    expect(applySpy).toHaveBeenCalledWith(
+      expect.objectContaining({ type: 'session.next.text.delta' })
+    );
+
+    // session.idle should close the stream and re-enable the input.
+    captured.onEvent({ type: 'session.idle', properties: { sessionID: 'ses-1' } });
+    expect(fakeSse.close).toHaveBeenCalled();
+    expect((prompt._sendBtn as HTMLButtonElement).disabled).toBe(false);
+  });
+
+  it('on submit: session.status with status.type=idle also closes the stream (idle shape variants)', async () => {
+    setOpenCodeProviders(null);
+    mockedCallOpenCodeEdit.mockResolvedValue({
+      ok: true,
+      sessionId: 'ses-1',
+      notebookPath: 'foo.ipynb'
+    });
+    const fakeSse = { close: jest.fn() };
+    let captured: any = null;
+    mockedSubscribeOpenCodeEvents.mockImplementation(
+      ((_sid: string, _s: any, handlers: any) => {
+        captured = handlers;
+        return fakeSse;
+      }) as any
+    );
+
+    const cell = makeFakeCell('x = 1');
+    const actions = new OpenCodeCellActions(cell);
+    Object.defineProperty(actions, 'parent', { value: cell, configurable: true });
+    (actions as any).onAfterAttach({} as any);
+    (actions.node.querySelector('button') as HTMLButtonElement).click();
+    const prompt = (actions as any)._prompt;
+    (prompt._textarea as HTMLTextAreaElement).value = 'go';
+    (prompt._sendBtn as HTMLButtonElement).click();
+    await new Promise(r => setTimeout(r, 5));
+
+    // session.status with status.type === 'idle' should also close the
+    // stream (some OpenCode versions emit this shape instead of
+    // session.idle).
+    captured.onEvent({
+      type: 'session.status',
+      properties: { sessionID: 'ses-1', status: { type: 'idle' } }
+    });
+    expect(fakeSse.close).toHaveBeenCalled();
+    expect((prompt._sendBtn as HTMLButtonElement).disabled).toBe(false);
+  });
+
+  it('on submit: idle nested under payload also closes the stream', async () => {
+    setOpenCodeProviders(null);
+    mockedCallOpenCodeEdit.mockResolvedValue({
+      ok: true,
+      sessionId: 'ses-1',
+      notebookPath: 'foo.ipynb'
+    });
+    const fakeSse = { close: jest.fn() };
+    let captured: any = null;
+    mockedSubscribeOpenCodeEvents.mockImplementation(
+      ((_sid: string, _s: any, handlers: any) => {
+        captured = handlers;
+        return fakeSse;
+      }) as any
+    );
+
+    const cell = makeFakeCell('x = 1');
+    const actions = new OpenCodeCellActions(cell);
+    Object.defineProperty(actions, 'parent', { value: cell, configurable: true });
+    (actions as any).onAfterAttach({} as any);
+    (actions.node.querySelector('button') as HTMLButtonElement).click();
+    const prompt = (actions as any)._prompt;
+    (prompt._textarea as HTMLTextAreaElement).value = 'go';
+    (prompt._sendBtn as HTMLButtonElement).click();
+    await new Promise(r => setTimeout(r, 5));
+
+    // Some OpenCode events have type at top level AND in payload.
+    captured.onEvent({
+      type: 'unknown.wrapper',
+      payload: {
+        type: 'session.idle',
+        properties: { sessionID: 'ses-1' }
+      }
+    });
+    expect(fakeSse.close).toHaveBeenCalled();
+    expect((prompt._sendBtn as HTMLButtonElement).disabled).toBe(false);
+  });
+
+  it('on submit failure: re-enables the input and shows an error notification', async () => {
+    setOpenCodeProviders(null);
+    mockedCallOpenCodeEdit.mockResolvedValue({
+      ok: false,
+      error: 'opencode down'
+    });
+
+    const cell = makeFakeCell('x = 1');
+    const actions = new OpenCodeCellActions(cell);
+    Object.defineProperty(actions, 'parent', { value: cell, configurable: true });
+    (actions as any).onAfterAttach({} as any);
+    const aiBtn = actions.node.querySelector('button') as HTMLButtonElement;
+    aiBtn.click();
+    const prompt = (actions as any)._prompt;
+    const textarea = prompt._textarea as HTMLTextAreaElement;
+    textarea.value = 'go';
+    (prompt._sendBtn as HTMLButtonElement).click();
+
+    await new Promise(r => setTimeout(r, 5));
+    expect(mockedNotification.error).toHaveBeenCalledWith('OpenCode 错误: opencode down');
+    expect((prompt._sendBtn as HTMLButtonElement).disabled).toBe(false);
   });
 });
 
@@ -674,4 +952,397 @@ describe('OpenCodeInlinePrompt', () => {
     );
     expect(onSubmit).not.toHaveBeenCalled();
   });
+
+  it('applyEvent: text deltas accumulate into a single assistant element (real-time streaming)', () => {
+    const cell = makeFakeCell('x = 1');
+    const prompt = new OpenCodeInlinePrompt(cell, {
+      onSubmit: jest.fn(),
+      onCancel: jest.fn(),
+      disabled: false,
+      providers: null
+    });
+    prompt.setSessionId('s1');
+
+    // Many deltas — they must all append to the SAME DOM node, not
+    // create a new one each time (that would defeat real-time rendering).
+    const deltas = ['hel', 'lo', ' ', 'wor', 'ld', '!', ' ', 'It works.'];
+    for (const d of deltas) {
+      prompt.applyEvent({
+        type: 'session.next.text.delta',
+        properties: { sessionID: 's1', delta: d }
+      });
+    }
+    const all = prompt.node.querySelectorAll(
+      '.opencode-inline-prompt .opencode-msg-assistant'
+    );
+    // Exactly one element — the streaming text is appended, not forked.
+    expect(all.length).toBe(1);
+    expect(all[0].textContent).toBe('hello world! It works.');
+  });
+
+  it('applyEvent: unrecognized event types are logged but never throw', () => {
+    const cell = makeFakeCell('x = 1');
+    const prompt = new OpenCodeInlinePrompt(cell, {
+      onSubmit: jest.fn(),
+      onCancel: jest.fn(),
+      disabled: false,
+      providers: null
+    });
+    const warnSpy = jest.spyOn(console, 'warn').mockImplementation(() => {});
+    prompt.setSessionId('s1');
+    // An event type the dispatcher doesn't recognize.
+    expect(() =>
+      prompt.applyEvent({
+        type: 'some.future.event',
+        properties: { sessionID: 's1', payload: 'whatever' }
+      })
+    ).not.toThrow();
+    expect(warnSpy).toHaveBeenCalledWith(
+      expect.stringContaining('unrecognized SSE event type'),
+      'some.future.event',
+      expect.anything()
+    );
+    warnSpy.mockRestore();
+  });
+
+  it('applyEvent: streaming markdown is rendered as HTML in real-time (not raw ```fence``` source)', () => {
+    // Regression: previously the streaming path appended deltas to
+    // .textContent, so the user saw raw ```py\n...\n``` source while
+    // the model was still typing. Only after closing+reopening the
+    // panel (which goes through setMessages + marked.parse) did the
+    // code block render as <pre><code>. The fix: re-render via
+    // marked.parse on every delta.
+    const cell = makeFakeCell('x = 1');
+    const prompt = new OpenCodeInlinePrompt(cell, {
+      onSubmit: jest.fn(),
+      onCancel: jest.fn(),
+      disabled: false,
+      providers: null
+    });
+    prompt.setSessionId('s1');
+    // Simulate a streaming reply that contains a fenced code block.
+    const deltas = [
+      'Here is the fix:\n\n',
+      '```python\n',
+      'import pandas as pd\n',
+      'import numpy as np\n',
+      '```\n'
+    ];
+    for (const d of deltas) {
+      prompt.applyEvent({
+        type: 'session.next.text.delta',
+        properties: { sessionID: 's1', delta: d }
+      });
+    }
+    // The <pre><code> block is rendered DURING streaming, not just at
+    // the end. The raw ```fence``` characters must NOT appear as visible
+    // text content in the message body.
+    const pre = prompt.node.querySelector(
+      '.opencode-msg-assistant .opencode-msg-content pre'
+    ) as HTMLElement;
+    expect(pre).not.toBeNull();
+    const code = pre.querySelector('code') as HTMLElement;
+    expect(code).not.toBeNull();
+    expect(code.textContent).toBe('import pandas as pd\nimport numpy as np');
+    // The Copy / Insert / Replace toolbar is attached to the <pre>
+    // block during streaming, same as in setMessages.
+    const toolbar = pre.parentElement!.querySelector(
+      '.opencode-code-toolbar'
+    ) as HTMLElement;
+    expect(toolbar).not.toBeNull();
+    expect(toolbar.querySelectorAll('button').length).toBe(3);
+    // The wrapper still has a single assistant element (not one per delta).
+    expect(
+      prompt.node.querySelectorAll(
+        '.opencode-inline-prompt .opencode-msg-assistant'
+      ).length
+    ).toBe(1);
+  });
+
+  it('applyEvent: message.part.delta with field=text routes to the assistant message', () => {
+    const cell = makeFakeCell('x = 1');
+    const prompt = new OpenCodeInlinePrompt(cell, {
+      onSubmit: jest.fn(),
+      onCancel: jest.fn(),
+      disabled: false,
+      providers: null
+    });
+    prompt.setSessionId('s1');
+    prompt.applyEvent({
+      type: 'message.part.delta',
+      properties: { sessionID: 's1', field: 'text', delta: 'world' }
+    });
+    const asst = prompt.node.querySelector(
+      '.opencode-inline-prompt .opencode-msg-assistant'
+    ) as HTMLElement;
+    expect(asst.textContent).toBe('world');
+  });
+
+  it('applyEvent: reasoning deltas render into a collapsible green block', () => {
+    const cell = makeFakeCell('x = 1');
+    const prompt = new OpenCodeInlinePrompt(cell, {
+      onSubmit: jest.fn(),
+      onCancel: jest.fn(),
+      disabled: false,
+      providers: null
+    });
+    prompt.setSessionId('s1');
+    prompt.applyEvent({
+      type: 'session.next.reasoning.delta',
+      properties: { sessionID: 's1', delta: 'thinking...' }
+    });
+    const block = prompt.node.querySelector(
+      '.opencode-inline-prompt .opencode-msg-block-reasoning'
+    ) as HTMLElement;
+    expect(block).not.toBeNull();
+    const content = block.querySelector(
+      '.opencode-msg-block-content'
+    ) as HTMLElement;
+    expect(content.textContent).toBe('thinking...');
+  });
+
+  it('applyEvent: tool-call lifecycle creates and updates a tool block', () => {
+    const cell = makeFakeCell('x = 1');
+    const prompt = new OpenCodeInlinePrompt(cell, {
+      onSubmit: jest.fn(),
+      onCancel: jest.fn(),
+      disabled: false,
+      providers: null
+    });
+    prompt.setSessionId('s1');
+    prompt.applyEvent({
+      type: 'session.next.tool.called',
+      properties: { sessionID: 's1', name: 'bash' }
+    });
+    prompt.applyEvent({
+      type: 'session.next.tool.input.delta',
+      properties: { sessionID: 's1', delta: '{"cmd":' }
+    });
+    prompt.applyEvent({
+      type: 'session.next.tool.success',
+      properties: { sessionID: 's1', result: { ok: true } }
+    });
+    const block = prompt.node.querySelector(
+      '.opencode-inline-prompt .opencode-msg-block-tool'
+    ) as HTMLElement;
+    expect(block).not.toBeNull();
+    const summary = block.querySelector('summary') as HTMLElement;
+    expect(summary.textContent).toContain('bash');
+    const content = block.querySelector(
+      '.opencode-msg-block-content'
+    ) as HTMLElement;
+    expect(content.textContent).toContain('{"cmd":');
+    expect(content.textContent).toContain('"ok": true');
+  });
+
+  it('applyEvent: session.idle resets stream pointers (next text delta starts a fresh assistant element)', () => {
+    const cell = makeFakeCell('x = 1');
+    const prompt = new OpenCodeInlinePrompt(cell, {
+      onSubmit: jest.fn(),
+      onCancel: jest.fn(),
+      disabled: false,
+      providers: null
+    });
+    prompt.setSessionId('s1');
+    prompt.applyEvent({
+      type: 'session.next.text.delta',
+      properties: { sessionID: 's1', delta: 'turn1' }
+    });
+    prompt.applyEvent({
+      type: 'session.idle',
+      properties: { sessionID: 's1' }
+    });
+    prompt.applyEvent({
+      type: 'session.next.text.delta',
+      properties: { sessionID: 's1', delta: 'turn2' }
+    });
+    const asstBlocks = prompt.node.querySelectorAll(
+      '.opencode-inline-prompt .opencode-msg-assistant'
+    );
+    expect(asstBlocks.length).toBe(2);
+    expect(asstBlocks[0].textContent).toBe('turn1');
+    expect(asstBlocks[1].textContent).toBe('turn2');
+  });
+
+  it('applyEvent: drops events for a different session (defensive filter)', () => {
+    const cell = makeFakeCell('x = 1');
+    const prompt = new OpenCodeInlinePrompt(cell, {
+      onSubmit: jest.fn(),
+      onCancel: jest.fn(),
+      disabled: false,
+      providers: null
+    });
+    prompt.setSessionId('s1');
+    prompt.applyEvent({
+      type: 'session.next.text.delta',
+      properties: { sessionID: 's2', delta: 'should be ignored' }
+    });
+    const asst = prompt.node.querySelector(
+      '.opencode-inline-prompt .opencode-msg-assistant'
+    );
+    expect(asst).toBeNull();
+  });
+
+  it('applyEvent: permission.asked creates a permission block with 3 buttons', () => {
+    const cell = makeFakeCell('x = 1');
+    const prompt = new OpenCodeInlinePrompt(cell, {
+      onSubmit: jest.fn(),
+      onCancel: jest.fn(),
+      disabled: false,
+      providers: null
+    });
+    prompt.setSessionId('s1');
+    prompt.applyEvent({
+      type: 'permission.asked',
+      properties: { sessionID: 's1', id: 'perm-1', description: 'rm -rf' }
+    });
+    const block = prompt.node.querySelector(
+      '.opencode-inline-prompt .opencode-msg-block-interaction'
+    ) as HTMLElement;
+    expect(block).not.toBeNull();
+    expect(block.id).toBe('opencode-perm-perm-1');
+    expect(block.textContent).toContain('rm -rf');
+    // Three buttons: once / always / reject.
+    const buttons = block.querySelectorAll('button');
+    expect(buttons.length).toBe(3);
+    expect(buttons[0].textContent).toContain('Once');
+    expect(buttons[1].textContent).toContain('Always');
+    expect(buttons[2].textContent).toContain('Reject');
+  });
+
+  it('applyEvent: clicking a permission button calls onPermissionReply and shows status', async () => {
+    const cell = makeFakeCell('x = 1');
+    const onPermissionReply = jest.fn().mockResolvedValue(true);
+    const prompt = new OpenCodeInlinePrompt(cell, {
+      onSubmit: jest.fn(),
+      onCancel: jest.fn(),
+      disabled: false,
+      providers: null,
+      onPermissionReply
+    });
+    prompt.setSessionId('s1');
+    prompt.applyEvent({
+      type: 'permission.asked',
+      properties: { sessionID: 's1', id: 'perm-9', description: 'sudo' }
+    });
+    const rejectBtn = prompt.node.querySelector(
+      '.opencode-perm-btn-reject'
+    ) as HTMLButtonElement;
+    rejectBtn.click();
+    await new Promise(r => setTimeout(r, 5));
+    expect(onPermissionReply).toHaveBeenCalledWith('perm-9', 'reject');
+    // The button group now shows the success status.
+    const block = prompt.node.querySelector(
+      '#opencode-perm-perm-9'
+    ) as HTMLElement;
+    expect(block.textContent).toContain('已拒绝');
+  });
+
+  it('applyEvent: question.asked creates a question block with input + submit', () => {
+    const cell = makeFakeCell('x = 1');
+    const prompt = new OpenCodeInlinePrompt(cell, {
+      onSubmit: jest.fn(),
+      onCancel: jest.fn(),
+      disabled: false,
+      providers: null
+    });
+    prompt.setSessionId('s1');
+    prompt.applyEvent({
+      type: 'question.asked',
+      properties: { sessionID: 's1', id: 'q-1', question: 'which env?' }
+    });
+    const block = prompt.node.querySelector(
+      '#opencode-q-q-1'
+    ) as HTMLElement;
+    expect(block).not.toBeNull();
+    expect(block.textContent).toContain('which env?');
+    expect(block.querySelector('input.opencode-q-input')).not.toBeNull();
+    expect(block.querySelector('button.opencode-q-submit')).not.toBeNull();
+  });
+
+  it('applyEvent: question submit calls onQuestionReply with the typed answer', async () => {
+    const cell = makeFakeCell('x = 1');
+    const onQuestionReply = jest.fn().mockResolvedValue(true);
+    const prompt = new OpenCodeInlinePrompt(cell, {
+      onSubmit: jest.fn(),
+      onCancel: jest.fn(),
+      disabled: false,
+      providers: null,
+      onQuestionReply
+    });
+    prompt.setSessionId('s1');
+    prompt.applyEvent({
+      type: 'question.asked',
+      properties: { sessionID: 's1', id: 'q-2', question: 'which version?' }
+    });
+    const input = prompt.node.querySelector(
+      '.opencode-q-input'
+    ) as HTMLInputElement;
+    const submit = prompt.node.querySelector(
+      '.opencode-q-submit'
+    ) as HTMLButtonElement;
+    input.value = 'use 3.12';
+    submit.click();
+    await new Promise(r => setTimeout(r, 5));
+    expect(onQuestionReply).toHaveBeenCalledWith('q-2', 'use 3.12');
+    const block = prompt.node.querySelector('#opencode-q-q-2') as HTMLElement;
+    expect(block.textContent).toContain('已回答');
+  });
+
+  it('applyEvent: question submit with empty input is a no-op', async () => {
+    const cell = makeFakeCell('x = 1');
+    const onQuestionReply = jest.fn();
+    const prompt = new OpenCodeInlinePrompt(cell, {
+      onSubmit: jest.fn(),
+      onCancel: jest.fn(),
+      disabled: false,
+      providers: null,
+      onQuestionReply
+    });
+    prompt.setSessionId('s1');
+    prompt.applyEvent({
+      type: 'question.asked',
+      properties: { sessionID: 's1', id: 'q-3', question: 'color?' }
+    });
+    const submit = prompt.node.querySelector(
+      '.opencode-q-submit'
+    ) as HTMLButtonElement;
+    submit.click();
+    await new Promise(r => setTimeout(r, 5));
+    expect(onQuestionReply).not.toHaveBeenCalled();
+  });
+
+  it('applyEvent: system events (server.* etc.) are NOT rendered', () => {
+    const cell = makeFakeCell('x = 1');
+    const prompt = new OpenCodeInlinePrompt(cell, {
+      onSubmit: jest.fn(),
+      onCancel: jest.fn(),
+      disabled: false,
+      providers: null
+    });
+    prompt.setSessionId('s1');
+    const systemTypes = [
+      'server.connected',
+      'workspace.updated',
+      'lsp.updated',
+      'mcp.tools.changed',
+      'pty.created',
+      'session.next.agent.switched',
+      'session.next.model.switched',
+      'file.edited'
+    ];
+    for (const t of systemTypes) {
+      prompt.applyEvent({ type: t, properties: { sessionID: 's1' } });
+    }
+    // No system lines, no message blocks, nothing rendered.
+    expect(
+      prompt.node.querySelector('.opencode-msg-system')
+    ).toBeNull();
+    expect(
+      prompt.node.querySelectorAll('.opencode-msg').length
+    ).toBe(0);
+    expect(
+      prompt.node.querySelectorAll('.opencode-msg-block').length
+    ).toBe(0);
+  });
 });

src/__tests__/opencode_client.spec.ts

@@ -1,6 +1,12 @@
 import { ServerConnection } from '@jupyterlab/services';
 
-import { callOpenCodeEdit, callOpenCodeProviders } from '../api/opencode_client';
+import {
+  callOpenCodeEdit,
+  callOpenCodeProviders,
+  callOpenCodeReplyPermission,
+  callOpenCodeReplyQuestion,
+  subscribeOpenCodeEvents
+} from '../api/opencode_client';
 import type { OpenCodeRequest } from '../types';
 
 jest.mock('@jupyterlab/services', () => {
@@ -82,10 +88,9 @@ describe('callOpenCodeEdit', () => {
     mockedMakeRequest.mockReset();
   });
 
-  it('POSTs to /opencode-bridge/edit with JSON body', async () => {
+  it('POSTs to /opencode-bridge/edit and returns async {ok, sessionId, notebookPath}', async () => {
     const respBody = {
       ok: true,
-      markdown: '```python\ndef foo(x: int) -> int: return x\n```',
       sessionId: 'sid',
       notebookPath: 'foo.ipynb',
     };
@@ -107,7 +112,10 @@ describe('callOpenCodeEdit', () => {
     );
     expect(resp.ok).toBe(true);
     if (resp.ok) {
-      expect(resp.markdown).toContain('int');
+      // Async response has no `markdown` field — the LLM reply arrives
+      // via the SSE stream, not embedded in this response.
+      expect(resp.sessionId).toBe('sid');
+      expect((resp as any).markdown).toBeUndefined();
     }
   });
 
@@ -219,3 +227,161 @@ describe('callOpenCodeProviders', () => {
     ).rejects.toThrow(/network error/);
   });
 });
+
+describe('callOpenCodeReplyPermission', () => {
+  beforeEach(() => {
+    mockedMakeRequest.mockReset();
+  });
+
+  it('POSTs to /opencode-bridge/permissions/:permId with {response}', async () => {
+    mockedMakeRequest.mockResolvedValue(
+      mockResponse({
+        ok: true,
+        status: 200,
+        statusText: 'OK',
+        text: JSON.stringify({ ok: true, permissionId: 'perm-1', response: 'once' })
+      })
+    );
+
+    const r = await callOpenCodeReplyPermission(
+      's1',
+      'perm-1',
+      'once',
+      mockServerSettings()
+    );
+    expect(r).toEqual({ ok: true, permissionId: 'perm-1', response: 'once' });
+    expect(mockedMakeRequest).toHaveBeenCalledWith(
+      'http://localhost:8888/opencode-bridge/permissions/perm-1?session=s1',
+      expect.objectContaining({
+        method: 'POST',
+        body: JSON.stringify({ response: 'once' })
+      }),
+      expect.anything()
+    );
+  });
+
+  it('throws on non-2xx', async () => {
+    mockedMakeRequest.mockResolvedValue(
+      mockResponse({
+        ok: false,
+        status: 400,
+        statusText: 'Bad Request',
+        text: 'bad response'
+      })
+    );
+    await expect(
+      callOpenCodeReplyPermission('s1', 'p', 'reject', mockServerSettings())
+    ).rejects.toThrow(/400/);
+  });
+});
+
+describe('callOpenCodeReplyQuestion', () => {
+  beforeEach(() => {
+    mockedMakeRequest.mockReset();
+  });
+
+  it('POSTs to /opencode-bridge/questions/:qId/reply with {answer}', async () => {
+    mockedMakeRequest.mockResolvedValue(
+      mockResponse({
+        ok: true,
+        status: 200,
+        statusText: 'OK',
+        text: JSON.stringify({ ok: true, questionId: 'q-1' })
+      })
+    );
+
+    const r = await callOpenCodeReplyQuestion(
+      's1',
+      'q-1',
+      'use 3.12',
+      mockServerSettings()
+    );
+    expect(r).toEqual({ ok: true, questionId: 'q-1' });
+    expect(mockedMakeRequest).toHaveBeenCalledWith(
+      'http://localhost:8888/opencode-bridge/questions/q-1/reply?session=s1',
+      expect.objectContaining({
+        method: 'POST',
+        body: JSON.stringify({ answer: 'use 3.12' })
+      }),
+      expect.anything()
+    );
+  });
+});
+
+/**
+ * subscribeOpenCodeEvents uses the global `fetch` (not ServerConnection),
+ * so we mock fetch instead of ServerConnection.makeRequest.
+ */
+describe('subscribeOpenCodeEvents', () => {
+  let originalFetch: typeof fetch;
+  let mockReader: { read: jest.Mock };
+  let capturedUrl: string | undefined;
+  let capturedInit: RequestInit | undefined;
+
+  beforeEach(() => {
+    originalFetch = globalThis.fetch;
+    mockReader = {
+      read: jest
+        .fn()
+        // First call returns a chunk; second returns done.
+        .mockResolvedValueOnce({
+          value: new TextEncoder().encode(
+            'data: {"type":"session.next.text.delta","properties":{"sessionID":"sid","delta":"hello"}}\n\n' +
+            'data: {"type":"session.idle","properties":{"sessionID":"sid"}}\n\n'
+          ),
+          done: false
+        })
+        .mockResolvedValueOnce({ value: undefined, done: true })
+    };
+    globalThis.fetch = jest.fn(async (url: any, init?: RequestInit) => {
+      capturedUrl = String(url);
+      capturedInit = init;
+      return {
+        ok: true,
+        status: 200,
+        statusText: 'OK',
+        body: {
+          getReader: () => mockReader
+        }
+      } as unknown as Response;
+    }) as unknown as typeof fetch;
+  });
+
+  afterEach(() => {
+    globalThis.fetch = originalFetch;
+  });
+
+  it('GETs /opencode-bridge/events?session=<sid> and dispatches parsed events', async () => {
+    const onEvent = jest.fn();
+    const settings = mockServerSettings();
+    const sub = subscribeOpenCodeEvents('sid', settings, { onEvent });
+
+    // Let the async fetch + reader resolve.
+    await new Promise(r => setTimeout(r, 10));
+    sub.close();
+
+    expect(capturedUrl).toBe(
+      'http://localhost:8888/opencode-bridge/events?session=sid'
+    );
+    expect(capturedInit?.method).toBe('GET');
+    expect((capturedInit?.headers as any).Accept).toBe('text/event-stream');
+    expect((capturedInit?.headers as any)['X-XSRFToken']).toBe('test');
+
+    expect(onEvent).toHaveBeenCalledTimes(2);
+    expect(onEvent.mock.calls[0][0].type).toBe('session.next.text.delta');
+    expect(onEvent.mock.calls[0][0].properties.delta).toBe('hello');
+    expect(onEvent.mock.calls[1][0].type).toBe('session.idle');
+  });
+
+  it('close() is idempotent and aborts the in-flight fetch', async () => {
+    const onEvent = jest.fn();
+    const sub = subscribeOpenCodeEvents('sid', mockServerSettings(), { onEvent });
+    // Let the async fetch actually run and capture the signal.
+    await new Promise(r => setTimeout(r, 5));
+    sub.close();
+    sub.close(); // second call must not throw
+    const signal = capturedInit?.signal as AbortSignal | undefined;
+    expect(signal).toBeDefined();
+    expect(signal?.aborted).toBe(true);
+  });
+});

src/api/opencode_client.ts

@@ -1,19 +1,32 @@
 /**
  * HTTP client for the opencode-bridge server extension.
+ *
+ * Two flows:
+ * - one-shot: POST /edit (async — returns immediately with sessionId),
+ *   GET /providers, GET /session-messages.
+ * - streaming: subscribeOpenCodeEvents() opens a fetch+reader connection
+ *   to GET /events and yields parsed OpenCodeEvent values.
  */
 import { URLExt } from '@jupyterlab/coreutils';
 import { ServerConnection } from '@jupyterlab/services';
 
-import type { OpenCodeProvidersResponse, OpenCodeRequest, OpenCodeResponse, OpenCodeMessagesResponse } from '../types';
+import type {
+  OpenCodeEditResponse,
+  OpenCodeEvent,
+  OpenCodeMessagesResponse,
+  OpenCodeProvidersResponse,
+  OpenCodeRequest
+} from '../types';
 
 /**
- * Call POST /opencode-bridge/edit. Returns parsed response (success or failure).
- * Throws on network error or non-JSON response.
+ * Call POST /opencode-bridge/edit. The endpoint is async — it fires
+ * prompt_async at OpenCode and returns once the prompt is queued. The
+ * actual LLM response is consumed via the SSE stream.
  */
 export async function callOpenCodeEdit(
   request: OpenCodeRequest,
   serverSettings: ServerConnection.ISettings
-): Promise<OpenCodeResponse> {
+): Promise<OpenCodeEditResponse> {
   const url = URLExt.join(
     serverSettings.baseUrl,
     'opencode-bridge',
@@ -23,7 +36,7 @@ export async function callOpenCodeEdit(
   const init: RequestInit = {
     method: 'POST',
     body: JSON.stringify(request),
-    headers: { 'Content-Type': 'application/json' },
+    headers: { 'Content-Type': 'application/json' }
   };
 
   let response: Response;
@@ -51,7 +64,7 @@ export async function callOpenCodeEdit(
     );
   }
 
-  return parsed as OpenCodeResponse;
+  return parsed as OpenCodeEditResponse;
 }
 
 /**
@@ -115,3 +128,233 @@ export async function callOpenCodeProviders(
 
   return (await response.json()) as OpenCodeProvidersResponse;
 }
+
+/**
+ * Respond to a `permission.asked` event. `response` must be one of
+ * "once" | "always" | "reject". Returns the parsed JSON body.
+ */
+export async function callOpenCodeReplyPermission(
+  sessionId: string,
+  permissionId: string,
+  response: 'once' | 'always' | 'reject',
+  serverSettings: ServerConnection.ISettings
+): Promise<{ ok: boolean; permissionId: string; response: string; error?: string }> {
+  const url =
+    URLExt.join(
+      serverSettings.baseUrl,
+      'opencode-bridge',
+      'permissions',
+      encodeURIComponent(permissionId)
+    ) +
+    '?session=' +
+    encodeURIComponent(sessionId);
+
+  const init: RequestInit = {
+    method: 'POST',
+    body: JSON.stringify({ response }),
+    headers: { 'Content-Type': 'application/json' }
+  };
+
+  let res: Response;
+  try {
+    res = await ServerConnection.makeRequest(url, init, serverSettings);
+  } catch (error) {
+    throw new Error(
+      `network error calling opencode-bridge/permissions: ${(error as Error).message}`
+    );
+  }
+  const text = await res.text();
+  if (!res.ok) {
+    throw new Error(
+      `opencode-bridge/permissions failed: ${res.status} ${text}`
+    );
+  }
+  return JSON.parse(text);
+}
+
+/**
+ * Reply to a `question.asked` event with the user's freeform answer.
+ */
+export async function callOpenCodeReplyQuestion(
+  sessionId: string,
+  questionId: string,
+  answer: string,
+  serverSettings: ServerConnection.ISettings
+): Promise<{ ok: boolean; questionId: string; error?: string }> {
+  const url =
+    URLExt.join(
+      serverSettings.baseUrl,
+      'opencode-bridge',
+      'questions',
+      encodeURIComponent(questionId),
+      'reply'
+    ) +
+    '?session=' +
+    encodeURIComponent(sessionId);
+
+  const init: RequestInit = {
+    method: 'POST',
+    body: JSON.stringify({ answer }),
+    headers: { 'Content-Type': 'application/json' }
+  };
+
+  let res: Response;
+  try {
+    res = await ServerConnection.makeRequest(url, init, serverSettings);
+  } catch (error) {
+    throw new Error(
+      `network error calling opencode-bridge/questions: ${(error as Error).message}`
+    );
+  }
+  const text = await res.text();
+  if (!res.ok) {
+    throw new Error(
+      `opencode-bridge/questions failed: ${res.status} ${text}`
+    );
+  }
+  return JSON.parse(text);
+}
+
+export interface OpenCodeEventHandlers {
+  onEvent: (event: OpenCodeEvent) => void;
+  onError?: (err: Error) => void;
+  onOpen?: () => void;
+}
+
+export interface OpenCodeEventSubscription {
+  /** Cancel the stream. Safe to call multiple times. */
+  close: () => void;
+}
+
+/**
+ * Subscribe to GET /opencode-bridge/events?session=<sid> as an SSE stream.
+ *
+ * Mirrors demo.html `initSSE`: opens a fetch(), reads the response body
+ * with a TextDecoder, splits on `\n\n` event boundaries, and parses each
+ * `data: {...}` line as JSON. Use `close()` to cancel.
+ *
+ * ServerConnection.makeRequest is unsuitable for SSE (it buffers the
+ * entire response), so we use plain fetch with the XSRF token injected
+ * from `serverSettings.token` when present.
+ */
+export function subscribeOpenCodeEvents(
+  sessionId: string,
+  serverSettings: ServerConnection.ISettings,
+  handlers: OpenCodeEventHandlers
+): OpenCodeEventSubscription {
+  const baseUrl = serverSettings.baseUrl.replace(/\/$/, '');
+  const url =
+    URLExt.join(baseUrl, 'opencode-bridge', 'events') +
+    '?session=' +
+    encodeURIComponent(sessionId);
+
+  const headers: Record<string, string> = {
+    Accept: 'text/event-stream'
+  };
+  if (serverSettings.token) {
+    headers['X-XSRFToken'] = serverSettings.token;
+  }
+
+  const controller = new AbortController();
+  let closed = false;
+
+  const close = (): void => {
+    if (closed) {
+      return;
+    }
+    closed = true;
+    try {
+      controller.abort();
+    } catch {
+      // ignore
+    }
+  };
+
+  void (async () => {
+    let response: Response;
+    try {
+      response = await fetch(url, { method: 'GET', headers, signal: controller.signal });
+    } catch (err) {
+      if ((err as Error).name === 'AbortError') {
+        return;
+      }
+      handlers.onError?.(err as Error);
+      return;
+    }
+    if (!response.ok) {
+      handlers.onError?.(
+        new Error(
+          `opencode-bridge/events failed: ${response.status} ${response.statusText}`
+        )
+      );
+      return;
+    }
+    if (!response.body) {
+      handlers.onError?.(new Error('opencode-bridge/events: no response body'));
+      return;
+    }
+    handlers.onOpen?.();
+
+    const reader = response.body.getReader();
+    const decoder = new TextDecoder('utf-8');
+    let buffer = '';
+
+    try {
+      while (!closed) {
+        const { value, done } = await reader.read();
+        if (done) {
+          break;
+        }
+        buffer += decoder.decode(value, { stream: true });
+        let sepIdx: number;
+        while ((sepIdx = buffer.indexOf('\n\n')) !== -1) {
+          const block = buffer.slice(0, sepIdx);
+          buffer = buffer.slice(sepIdx + 2);
+          for (const line of block.split('\n')) {
+            if (!line.startsWith('data:')) {
+              continue;
+            }
+            const jsonStr = line.slice(5).trim();
+            if (!jsonStr) {
+              continue;
+            }
+            try {
+              const event = JSON.parse(jsonStr) as OpenCodeEvent;
+              handlers.onEvent(event);
+            } catch (e) {
+              console.warn(
+                'opencode_bridge: SSE JSON parse error:',
+                e,
+                jsonStr
+              );
+            }
+          }
+        }
+      }
+      // Drain trailing block (no final \n\n).
+      if (buffer.trim()) {
+        for (const line of buffer.split('\n')) {
+          if (!line.startsWith('data:')) {
+            continue;
+          }
+          const jsonStr = line.slice(5).trim();
+          if (!jsonStr) {
+            continue;
+          }
+          try {
+            handlers.onEvent(JSON.parse(jsonStr) as OpenCodeEvent);
+          } catch {
+            // ignore
+          }
+        }
+      }
+    } catch (err) {
+      if ((err as Error).name === 'AbortError') {
+        return;
+      }
+      handlers.onError?.(err as Error);
+    }
+  })();
+
+  return { close };
+}

src/components/model_selection.ts

@@ -0,0 +1,82 @@
+/**
+ * Persistence of the user's (provider, model) selection in the inline
+ * prompt, keyed by notebook path. Stored in localStorage so the
+ * preference survives panel close/reopen and JupyterLab restart.
+ *
+ * The selection is ALWAYS validated against the current providers
+ * payload before being applied: if the stored provider (or the model
+ * under it) is no longer present, the caller falls back to its
+ * default-selection logic. We never silently apply a stale value that
+ * would result in an empty / disabled <select>.
+ *
+ * Storage is best-effort: localStorage may be disabled (private mode),
+ * the quota may be exhausted, or the value may be corrupt. All
+ * failure modes return null / no-op rather than throw.
+ */
+const STORAGE_PREFIX = 'opencode_bridge:model-selection:';
+
+export interface ModelSelection {
+  providerId: string;
+  modelId: string;
+}
+
+export function loadModelSelection(notebookPath: string): ModelSelection | null {
+  if (!notebookPath) {
+    return null;
+  }
+  try {
+    if (typeof localStorage === 'undefined') {
+      return null;
+    }
+    const raw = localStorage.getItem(STORAGE_PREFIX + notebookPath);
+    if (!raw) {
+      return null;
+    }
+    const parsed = JSON.parse(raw);
+    if (
+      parsed &&
+      typeof parsed === 'object' &&
+      typeof (parsed as ModelSelection).providerId === 'string' &&
+      typeof (parsed as ModelSelection).modelId === 'string'
+    ) {
+      return {
+        providerId: (parsed as ModelSelection).providerId,
+        modelId: (parsed as ModelSelection).modelId
+      };
+    }
+    return null;
+  } catch {
+    return null;
+  }
+}
+
+export function saveModelSelection(
+  notebookPath: string,
+  sel: ModelSelection
+): void {
+  if (!notebookPath) {
+    return;
+  }
+  try {
+    if (typeof localStorage === 'undefined') {
+      return;
+    }
+    localStorage.setItem(STORAGE_PREFIX + notebookPath, JSON.stringify(sel));
+  } catch {
+    // Ignore quota exceeded / disabled storage.
+  }
+}
+
+export function clearModelSelection(notebookPath: string): void {
+  if (!notebookPath) {
+    return;
+  }
+  try {
+    if (typeof localStorage === 'undefined') {
+      return;
+    }
+    localStorage.removeItem(STORAGE_PREFIX + notebookPath);
+  } catch {
+    // ignore
+  }
+}

src/components/opencode_cell_actions.ts

@@ -1,14 +1,20 @@
 /**
  * Per-cell AI action button rendered inside JupyterLab's native Cell toolbar
- * (top-right of the active cell). v2: a single icon button. Clicking it
- * toggles an OpenCodeInlinePrompt panel attached to the cell's DOM.
+ * (top-right of the active cell). Clicking it toggles an
+ * OpenCodeInlinePrompt panel attached to the cell's DOM.
  *
  * Registered in index.ts via IToolbarWidgetRegistry.addFactory('Cell', ...).
  *
- * v3-final: the frontend does NO semantic processing. It gathers the cell's
- * source / outputs / error, attaches the user's freeform instruction, and
- * POSTs to the server. The chosen provider/model comes from the inline
- * picker (no settings-based default).
+ * v4 message flow (matches demo.html):
+ *   1. User clicks 🪄 -> panel appears, shows session history (setMessages)
+ *   2. User submits prompt -> callOpenCodeEdit returns immediately with
+ *      sessionId
+ *   3. We subscribe to GET /opencode-bridge/events?session=<sid> and
+ *      forward each event to the prompt's applyEvent() dispatcher
+ *   4. On session.idle we close the stream and re-enable the input
+ *
+ * The frontend does NO semantic processing of the LLM reply. The chosen
+ * provider/model comes from the inline picker (no settings-based default).
  */
 import type { CodeCell } from '@jupyterlab/cells';
 import { Notification } from '@jupyterlab/apputils';
@@ -16,13 +22,18 @@ import type { NotebookPanel } from '@jupyterlab/notebook';
 import { ServerConnection } from '@jupyterlab/services';
 import { Widget } from '@lumino/widgets';
 
-import { callOpenCodeEdit, callOpenCodeSessionMessages } from '../api/opencode_client';
+import {
+  callOpenCodeEdit,
+  callOpenCodeReplyPermission,
+  callOpenCodeReplyQuestion,
+  callOpenCodeSessionMessages,
+  subscribeOpenCodeEvents,
+  type OpenCodeEventSubscription
+} from '../api/opencode_client';
 import { extractCellContext } from '../context/cell_context';
 import type {
   CellContext,
-  OpenCodeProvidersResponse,
-  OpenCodeRequest,
-  OpenCodeResponse
+  OpenCodeProvidersResponse
 } from '../types';
 
 import { OpenCodeInlinePrompt } from './opencode_inline_prompt';
@@ -50,6 +61,7 @@ export class OpenCodeCellActions extends Widget {
   private _context: CellContext | null = null;
   private _status: Status = 'idle';
   private _prompt: OpenCodeInlinePrompt | null = null;
+  private _sseSub: OpenCodeEventSubscription | null = null;
 
   constructor(cell: CodeCell) {
     super();
@@ -65,6 +77,7 @@ export class OpenCodeCellActions extends Widget {
     if (this.isDisposed) {
       return;
     }
+    this._closeSse();
     this._cell.model.contentChanged.disconnect(this._onModelChange, this);
     this._cell.model.stateChanged.disconnect(this._onModelChange, this);
     this._hidePrompt();
@@ -114,19 +127,67 @@ export class OpenCodeCellActions extends Widget {
     const prompt = new OpenCodeInlinePrompt(this._cell, {
       disabled: !this._context || this._status === 'loading',
       providers: _providers,
+      notebookPath: this._context?.notebookPath,
       onSubmit: (text: string, providerId?: string, modelId?: string) => {
         void this._onSubmit(text, providerId, modelId);
       },
       onCancel: () => {
         this._hidePrompt();
+      },
+      onPermissionReply: async (permId, response) => {
+        if (!_serverSettings) {
+          return false;
+        }
+        const sessionId = (this._prompt as any)._sessionId as string | null;
+        if (!sessionId) {
+          return false;
+        }
+        try {
+          const r = await callOpenCodeReplyPermission(
+            sessionId,
+            permId,
+            response,
+            _serverSettings
+          );
+          return r.ok;
+        } catch (e) {
+          console.warn('opencode_bridge: permission reply failed', e);
+          return false;
+        }
+      },
+      onQuestionReply: async (qId, answer) => {
+        if (!_serverSettings) {
+          return false;
+        }
+        const sessionId = (this._prompt as any)._sessionId as string | null;
+        if (!sessionId) {
+          return false;
+        }
+        try {
+          const r = await callOpenCodeReplyQuestion(
+            sessionId,
+            qId,
+            answer,
+            _serverSettings
+          );
+          return r.ok;
+        } catch (e) {
+          console.warn('opencode_bridge: question reply failed', e);
+          return false;
+        }
+      },
+      onStreamEnd: () => {
+        // Centralized: the prompt detected "idle" (any shape) and tells
+        // us to close the SSE stream + restore the input. We no longer
+        // parse event types here.
+        this._closeSse();
       }
     });
     Widget.attach(prompt, this._cell.node);
     this._prompt = prompt;
 
-    // Fetch the current session's message history and render it into
-    // the prompt's scrollable history area. Empty list (no session yet)
-    // is a normal no-op render.
+    // Fetch the current session's static history and render it before
+    // streaming begins. Empty list is a normal no-op.
     const notebookPath = this._context?.notebookPath;
     if (notebookPath) {
       void this._refreshHistory(notebookPath);
@@ -144,8 +205,6 @@ export class OpenCodeCellActions extends Widget {
       );
       this._prompt.setMessages(resp.messages);
     } catch (e) {
-      // Non-fatal: the history is a convenience. The user can still
-      // send new prompts; just the scrollback won't update.
       console.warn('opencode_bridge: failed to fetch session messages', e);
     }
   }
@@ -170,7 +229,7 @@ export class OpenCodeCellActions extends Widget {
       return;
     }
 
-    const request: OpenCodeRequest = {
+    const request = {
       prompt: text,
       context: this._context,
       providerId,
@@ -180,41 +239,55 @@ export class OpenCodeCellActions extends Widget {
     this._status = 'loading';
     if (this._prompt) {
       this._prompt.setDisabled(true);
+      this._prompt.setSessionId(null);
     }
 
+    let resp;
     try {
-      const resp = await callOpenCodeEdit(request, _serverSettings);
-      this._handleResponse(resp);
+      resp = await callOpenCodeEdit(request, _serverSettings);
     } catch (e) {
       this._status = 'idle';
-      if (this._prompt) {
-        this._prompt.setDisabled(false);
-      }
+      this._prompt?.setDisabled(false);
       Notification.error(`OpenCode 失败: ${(e as Error).message}`);
+      return;
     }
+
+    if (!resp.ok) {
+      this._status = 'idle';
+      this._prompt?.setDisabled(false);
+      Notification.error(`OpenCode 错误: ${resp.error}`);
+      return;
+    }
+
+    // Async edit accepted. Bind the prompt to the returned session, clear
+    // the input, then subscribe to the SSE event stream for this session.
+    this._prompt?.setSessionId(resp.sessionId);
+    this._prompt?.clearInput();
+    this._sseSub = subscribeOpenCodeEvents(
+      resp.sessionId,
+      _serverSettings,
+      {
+        onEvent: ev => {
+          this._prompt?.applyEvent(ev);
+          // Stream-end is detected by the prompt and signalled via
+          // onStreamEnd (see _showPrompt); no idle parsing here.
+        },
+        onError: err => {
+          console.warn('opencode_bridge: SSE error', err);
+          // Don't flip status here: the server may just have hiccuped.
+          // session.idle (if it comes through) will re-enable the input.
+        }
+      }
+    );
   }
 
-  private _handleResponse(resp: OpenCodeResponse): void {
-    this._status = 'idle';
-    if (resp.ok) {
-      // Refetch the (now-updated) session history and render the new
-      // assistant message as the last item in the scrollable history.
-      // The cell source is NOT replaced.
-      const notebookPath = this._context?.notebookPath;
-      if (notebookPath) {
-        void this._refreshHistory(notebookPath);
-      }
-      // Clear the input so the user can type a follow-up message.
-      this._prompt?.clearInput();
-      Notification.info(
-        `OpenCode 完成 (${resp.markdown.length} chars). Session: ${resp.sessionId.slice(0, 8)}…`
-      );
-    } else {
-      if (this._prompt) {
-        this._prompt.setDisabled(false);
-      }
-      Notification.error(`OpenCode 错误: ${resp.error}`);
+  private _closeSse(): void {
+    if (this._sseSub) {
+      this._sseSub.close();
+      this._sseSub = null;
     }
+    this._status = 'idle';
+    this._prompt?.setDisabled(false);
   }
 }
 

src/components/opencode_inline_prompt.ts

@@ -1,26 +1,70 @@
 /**
  * Inline prompt widget attached to a cell's DOM when the AI button is clicked.
- * Renders two <select> boxes (provider + model), a textarea, send/cancel
- * buttons, and a scrollable history area that shows the current session's
- * messages (user + assistant). The model select's default for the chosen
- * provider is `default[providerID]` from the /config/providers response.
  *
- * v3-final: no settings — picks come from the OpenCode Serve providers
- * cache; history comes from /session-messages; cell source is NOT replaced
- * (the AI reply is rendered as markdown inside this history area).
+ * Renders two <select> boxes (provider + model), a textarea, send/cancel
+ * buttons, and a scrollable history area. The history area serves two modes:
+ *
+ * 1. **Static** — `setMessages(messages)` renders the session's stored
+ *    history (user/assistant markdown) when the panel is first opened.
+ * 2. **Streaming** — after the user submits a prompt, the owning cell
+ *    action subscribes to `/opencode-bridge/events` and forwards every
+ *    `OpenCodeEvent` to `applyEvent(event)`, which dispatches to the
+ *    appropriate DOM helper (text deltas, reasoning blocks, tool blocks,
+ *    permission/question prompts, system lines). On `session.idle` the
+ *    stream pointers reset.
+ *
+ * The model select's default for the chosen provider is `default[providerID]`
+ * from the /config/providers response.
+ *
+ * v4: cell source is NOT replaced. The AI reply is rendered into this
+ * history area. The user can still Copy / Insert / Replace individual
+ * code blocks from a finished assistant message.
  */
 import { CodeCell } from '@jupyterlab/cells';
 import { Notification } from '@jupyterlab/apputils';
 import { Widget } from '@lumino/widgets';
 import { marked } from 'marked';
 
-import type { OpenCodeMessage, OpenCodeProvidersResponse } from '../types';
+import type {
+  OpenCodeEvent,
+  OpenCodeMessage,
+  OpenCodeProvidersResponse
+} from '../types';
+import {
+  loadModelSelection,
+  saveModelSelection
+} from './model_selection';
 
 export interface IOpenCodeInlinePromptOptions {
   onSubmit: (text: string, providerId?: string, modelId?: string) => void;
   onCancel: () => void;
   disabled: boolean;
   providers: OpenCodeProvidersResponse | null;
+  /**
+   * Notebook path used to key the persisted (provider, model) selection
+   * in localStorage. If omitted (or the stored selection is no longer
+   * valid in the current `providers` payload), the prompt falls back
+   * to the default selection (first provider / `default[pid]` model).
+   */
+  notebookPath?: string;
+  /**
+   * Respond to a `permission.asked` event. Returns a promise that resolves
+   * to true on success, false on failure (the prompt updates the UI in
+   * both cases).
+   */
+  onPermissionReply?: (
+    permissionId: string,
+    response: 'once' | 'always' | 'reject'
+  ) => Promise<boolean>;
+  /** Reply to a `question.asked` event with the user's freeform answer. */
+  onQuestionReply?: (questionId: string, answer: string) => Promise<boolean>;
+  /**
+   * Fired when the prompt detects the agent has finished its turn
+   * (any of: `session.idle`, `session.status` with `status.type === 'idle'`,
+   * or the same shapes nested under `payload`). Centralizing this in
+   * the prompt keeps the cell action from re-parsing event types.
+   */
+  onStreamEnd?: () => void;
 }
 
 interface FlatProvider {
@@ -65,8 +109,14 @@ function pickDefaultModelId(
   return keys[0];
 }
 
+interface BlockEntry {
+  details: HTMLDetailsElement;
+  content: HTMLDivElement;
+}
+
 export class OpenCodeInlinePrompt extends Widget {
   private _cell: CodeCell;
+  private _options: IOpenCodeInlinePromptOptions;
   private _textarea: HTMLTextAreaElement;
   private _sendBtn: HTMLButtonElement;
   private _cancelBtn: HTMLButtonElement;
@@ -76,18 +126,40 @@ export class OpenCodeInlinePrompt extends Widget {
   private _defaultMap: { [pid: string]: string };
   private _historyEl: HTMLDivElement;
 
+  // Streaming state. _sessionId is set by the cell action after a successful
+  // /edit response; applyEvent() drops events that don't belong to it.
+  private _sessionId: string | null = null;
+  // The current assistant message DOM node (re-rendered as markdown on
+  // every text delta so the user sees a properly formatted response in
+  // real-time, not raw ```fence``` source).
+  private _currentAssistantEl: HTMLDivElement | null = null;
+  // Accumulator for the current assistant turn's raw markdown source.
+  // Each text delta appends here; the whole string is then re-rendered
+  // through marked.parse to refresh the DOM. Using the raw source as
+  // truth (not innerHTML) avoids lossy round-trips through partial HTML.
+  private _currentAssistantText: string = '';
+  // Active collapsible blocks (reasoning / tool / etc.), keyed by type.
+  private _activeBlocks: { [key: string]: BlockEntry } = {};
+
   constructor(
     cell: CodeCell,
     options: IOpenCodeInlinePromptOptions
   ) {
     super();
     this._cell = cell;
+    this._options = options;
     this.addClass('opencode-inline-prompt');
 
     const flat = flattenProviders(options.providers);
     this._providerModels = flat.providers;
     this._defaultMap = flat.defaultMap;
 
+    // Resolve the initial selection: prefer a previously-saved
+    // (providerId, modelId) IF it's still valid in the current
+    // providers payload; otherwise fall back to the default
+    // (first provider / `default[pid]` model).
+    const stored = this._resolveStoredSelection(options.notebookPath);
+
     if (this._providerModels.length > 0) {
       this._providerSelect = document.createElement('select');
       this._providerSelect.className = 'opencode-provider-select';
@@ -97,16 +169,25 @@ export class OpenCodeInlinePrompt extends Widget {
         o.textContent = p.name ? `${p.id} (${p.name})` : p.id;
         this._providerSelect.appendChild(o);
       }
-      this._providerSelect.value = this._providerModels[0].id;
+      this._providerSelect.value = stored
+        ? stored.providerId
+        : this._providerModels[0].id;
       this._providerSelect.addEventListener('change', () => {
         this._rebuildModelSelect(this._providerSelect!.value);
+        this._persistSelection();
       });
     }
 
     if (this._providerModels.length > 0) {
       this._modelSelect = document.createElement('select');
       this._modelSelect.className = 'opencode-model-select';
-      this._rebuildModelSelect(this._providerSelect?.value);
+      this._rebuildModelSelect(
+        this._providerSelect?.value,
+        stored ? stored.modelId : undefined
+      );
+      this._modelSelect.addEventListener('change', () => {
+        this._persistSelection();
+      });
     }
 
     this._textarea = document.createElement('textarea');
@@ -125,8 +206,6 @@ export class OpenCodeInlinePrompt extends Widget {
       const modelId = this._modelSelect?.value || undefined;
       options.onSubmit(text, providerId, modelId);
     });
-    // Chat-input convention: Enter submits, Shift+Enter / Ctrl+Enter
-    // insert a newline (textarea default).
     this._textarea.addEventListener('keydown', (e) => {
       if (e.key === 'Enter' && !e.shiftKey && !e.ctrlKey) {
         e.preventDefault();
@@ -167,8 +246,8 @@ export class OpenCodeInlinePrompt extends Widget {
       this.node.appendChild(row);
     }
 
-    // Scrollable history of the current session's messages. Empty until
-    // setMessages() is called by the owning cell action.
+    // Scrollable history (static + streaming). Empty until setMessages() /
+    // applyEvent() populate it.
     this._historyEl = document.createElement('div');
     this._historyEl.className = 'opencode-inline-history';
     this.node.appendChild(this._historyEl);
@@ -177,7 +256,10 @@ export class OpenCodeInlinePrompt extends Widget {
     this.node.appendChild(actions);
   }
 
-  private _rebuildModelSelect(providerId: string | undefined): void {
+  private _rebuildModelSelect(
+    providerId: string | undefined,
+    preferredModelId?: string
+  ): void {
     if (!this._modelSelect) {
       return;
     }
@@ -205,8 +287,65 @@ export class OpenCodeInlinePrompt extends Widget {
       this._modelSelect.appendChild(o);
     }
     this._modelSelect.disabled = false;
-    this._modelSelect.value =
-      pickDefaultModelId(providerId, provider.models, this._defaultMap) ?? keys[0];
+    // Honor a preferred model (e.g. from a stored selection), but only
+    // if it actually exists in this provider's model list. Otherwise
+    // fall back to the default.
+    if (
+      preferredModelId &&
+      Object.prototype.hasOwnProperty.call(provider.models, preferredModelId)
+    ) {
+      this._modelSelect.value = preferredModelId;
+    } else {
+      this._modelSelect.value =
+        pickDefaultModelId(providerId, provider.models, this._defaultMap) ??
+        keys[0];
+    }
+  }
+
+  /**
+   * Read the stored (providerId, modelId) for this notebook and
+   * validate it against the current providers payload. Returns the
+   * stored selection only when BOTH the provider and the model still
+   * exist; otherwise returns null and the caller uses the default.
+   */
+  private _resolveStoredSelection(
+    notebookPath: string | undefined
+  ): { providerId: string; modelId: string } | null {
+    if (!notebookPath) {
+      return null;
+    }
+    const sel = loadModelSelection(notebookPath);
+    if (!sel) {
+      return null;
+    }
+    const provider = this._providerModels.find(p => p.id === sel.providerId);
+    if (!provider) {
+      // The provider was removed since the user last picked.
+      return null;
+    }
+    if (!Object.prototype.hasOwnProperty.call(provider.models, sel.modelId)) {
+      // The model was removed (or renamed) under a still-valid provider.
+      return null;
+    }
+    return sel;
+  }
+
+  /**
+   * Persist the current (providerSelect.value, modelSelect.value) for
+   * this notebook. No-op if the prompt has no notebookPath, no
+   * selects, or the values are empty.
+   */
+  private _persistSelection(): void {
+    const notebookPath = this._options.notebookPath;
+    if (!notebookPath || !this._providerSelect || !this._modelSelect) {
+      return;
+    }
+    const providerId = this._providerSelect.value;
+    const modelId = this._modelSelect.value;
+    if (!providerId || !modelId) {
+      return;
+    }
+    saveModelSelection(notebookPath, { providerId, modelId });
   }
 
   setDisabled(disabled: boolean): void {
@@ -219,20 +358,27 @@ export class OpenCodeInlinePrompt extends Widget {
     }
   }
 
-  /**
-   * Clear the user input textarea (called by the cell action after a
-   * successful response, so the user can type a follow-up message
-   * without manually deleting the previous prompt).
-   */
+  /** Clear the user input textarea. */
   clearInput(): void {
     this._textarea.value = '';
   }
 
   /**
-   * Render the current session's messages into the scrollable history
-   * area. User messages are plain text; assistant messages are rendered
-   * as markdown via marked. Auto-scrolls to the bottom so the most
-   * recent message is visible.
+   * Bind the prompt to a specific OpenCode session id. Events arriving via
+   * `applyEvent` for a different session are dropped. Pass null to clear.
+   */
+  setSessionId(sessionId: string | null): void {
+    this._sessionId = sessionId;
+    if (sessionId) {
+      // Fresh turn: drop any in-progress stream DOM from a previous session.
+      this._resetStreamPointers();
+    }
+  }
+
+  /**
+   * Render the current session's static messages (loaded once from
+   * /session-messages when the panel opens). Subsequent assistant turns
+   * during the same session are appended live via applyEvent().
    */
   setMessages(messages: OpenCodeMessage[]): void {
     this._historyEl.textContent = '';
@@ -242,74 +388,501 @@ export class OpenCodeInlinePrompt extends Widget {
       if (msg.role === 'user') {
         wrap.textContent = msg.content;
       } else {
-        // Assistant: render the markdown, then for every fenced code
-        // block (<pre>) attach a small toolbar with Copy / Insert /
-        // Replace buttons that act on THAT block's code (not the whole
-        // message). Messages without code fences get no toolbar.
-        const rendered = document.createElement('div');
-        rendered.className = 'opencode-msg-content';
-        rendered.innerHTML = marked.parse(msg.content) as string;
-
-        const codeBlocks = rendered.querySelectorAll('pre');
-        codeBlocks.forEach(pre => {
-          const code = pre.querySelector('code');
-          if (!code) {
-            return;
-          }
-          // Snapshot the code text before we mutate the DOM so the
-          // buttons act on the exact code inside this block (no
-          // surrounding ```fences```, no toolbar label text).
-          const codeText = code.textContent ?? '';
-          const wrap = document.createElement('div');
-          wrap.className = 'opencode-code-block';
-          const toolbar = document.createElement('div');
-          toolbar.className = 'opencode-code-toolbar';
-          const mkBtn = (
-            label: string,
-            title: string,
-            onClick: () => void
-          ) => {
-            const b = document.createElement('button');
-            b.type = 'button';
-            b.className = 'opencode-code-btn';
-            b.textContent = label;
-            b.title = title;
-            b.addEventListener('click', onClick);
-            return b;
-          };
-          toolbar.appendChild(
-            mkBtn('📋 复制', '复制代码', () => {
-              void this._copyToClipboard(codeText);
-            })
-          );
-          toolbar.appendChild(
-            mkBtn('↪ 插入', '在光标位置插入', () => {
-              this._insertAtCursor(codeText);
-            })
-          );
-          toolbar.appendChild(
-            mkBtn('⟳ 替换', '替换单元格内容', () => {
-              this._replaceCell(codeText);
-            })
-          );
-          wrap.appendChild(toolbar);
-          // Move the original <pre> into the wrapper.
-          pre.parentNode!.insertBefore(wrap, pre);
-          wrap.appendChild(pre);
-        });
-
-        wrap.appendChild(rendered);
+        this._renderAssistantMarkdown(wrap, msg.content);
       }
       this._historyEl.appendChild(wrap);
     }
-    this._historyEl.scrollTop = this._historyEl.scrollHeight;
+    this._scrollToBottom();
   }
 
   /**
-   * Copy the given text to the clipboard. Prefers navigator.clipboard
-   * (available in secure contexts incl. JupyterLab). Falls back to a
-   * hidden textarea + execCommand for older / non-secure contexts.
+   * Apply one OpenCode SSE event to the streaming UI. Mirrors
+   * demo.html's handleGlobalEvent / processSessionEvent dispatch.
    */
+  applyEvent(event: OpenCodeEvent): void {
+    const type = event.type || (event.payload && event.payload.type) || '';
+    const props =
+      event.properties ||
+      (event.payload && event.payload.properties) ||
+      {};
+    const sessionID =
+      props.sessionID ||
+      props.sessionId ||
+      (event.syncEvent && event.syncEvent.aggregateID) ||
+      (event.payload &&
+        event.payload.syncEvent &&
+        event.payload.syncEvent.aggregateID);
+
+    // Drop events for other sessions (defensive; matches demo.html
+    // handleGlobalEvent). Only drop when the event has an explicit
+    // sessionID — events without one (e.g. system events) are still
+    // processed and may be dropped by _isSystemEvent below.
+    if (sessionID && this._sessionId && sessionID !== this._sessionId) {
+      return;
+    }
+
+    // System / workspace / pty / etc. events are not displayed.
+    if (this._isSystemEvent(type)) {
+      return;
+    }
+
+    // For text/reasoning/tool/idle/permission/question we expect to match.
+    // If none of the branches below fire, log a warning so the missing
+    // event type can be diagnosed in DevTools.
+    const handled = this._dispatchContentEvent(event, type, props);
+    if (!handled) {
+      console.warn(
+        'opencode_bridge: unrecognized SSE event type (dropped):',
+        type,
+        event
+      );
+    }
+  }
+
+  private _dispatchContentEvent(
+    event: OpenCodeEvent,
+    type: string,
+    props: { [k: string]: any }
+  ): boolean {
+
+    if (type === 'permission.asked') {
+      const permId = String(props.id || props.permissionID || '');
+      this._createPermissionBlock(
+        permId,
+        String(
+          props.description || props.title || '系统请求执行权限（命令行/文件修改）'
+        )
+      );
+      return true;
+    }
+    if (type === 'question.asked') {
+      const qId = String(props.id || props.questionID || '');
+      this._createQuestionBlock(
+        qId,
+        String(props.question || props.title || 'Agent 正在等待确认…')
+      );
+      return true;
+    }
+
+    if (
+      type === 'message.part.delta' &&
+      props.field === 'text' &&
+      typeof props.delta === 'string'
+    ) {
+      this._appendToAssistantText(props.delta);
+      return true;
+    }
+    if (
+      (type === 'message.part.delta' && props.field === 'reasoning') ||
+      type === 'session.next.reasoning.delta'
+    ) {
+      this._appendToBlock(
+        'reasoning',
+        '🧠 思考与推理过程',
+        String(props.delta || props.text || '')
+      );
+      return true;
+    }
+    if (type === 'session.next.text.delta' && typeof props.delta === 'string') {
+      this._appendToAssistantText(props.delta);
+      return true;
+    }
+
+    if (type === 'session.next.reasoning.started') {
+      this._getOrCreateBlock('reasoning', '🧠 思考与推理过程');
+      return true;
+    }
+
+    if (
+      type === 'session.next.tool.input.started' ||
+      type === 'session.next.tool.called'
+    ) {
+      const toolName = String(props.name || props.tool || '未知工具');
+      this._getOrCreateBlock('tool', `🛠️ 调用工具: ${toolName}`);
+      return true;
+    }
+    if (type === 'session.next.tool.input.delta') {
+      this._appendToBlock('tool', '🛠️ 工具输入与参数', String(props.delta || ''));
+      return true;
+    }
+    if (type === 'session.next.tool.progress') {
+      this._appendToBlock(
+        'tool',
+        '🛠️ 工具执行中…',
+        `\n[Progress]: ${String(props.message || '')}`
+      );
+      return true;
+    }
+    if (type === 'session.next.tool.success') {
+      this._appendToBlock(
+        'tool',
+        '🛠️ 工具执行完毕',
+        `\n[执行成功结果]: ${JSON.stringify(props.result || {}, null, 2)}`
+      );
+      return true;
+    }
+    if (type === 'session.next.tool.failed') {
+      this._appendToBlock(
+        'tool',
+        '🛠️ 工具执行失败',
+        `\n[执行失败异常]: ${String(props.error || 'Unknown')}`
+      );
+      return true;
+    }
+
+    if (this._isIdleEvent(event, type, props)) {
+      this._resetStreamPointers();
+      this._options.onStreamEnd?.();
+      return true;
+    }
+
+    return false;
+  }
+
+  /**
+   * Detect "agent is now idle" across the multiple shapes OpenCode has
+   * emitted across versions: top-level `session.idle`, top-level
+   * `session.status` with `status.type === 'idle'`, and the same shapes
+   * nested under `payload`.
+   */
+  private _isIdleEvent(
+    event: OpenCodeEvent,
+    topType: string,
+    topProps: { [k: string]: any }
+  ): boolean {
+    const check = (t: string, p: { [k: string]: any }): boolean => {
+      if (t === 'session.idle') {
+        return true;
+      }
+      if (t === 'session.status' && p && p.status && p.status.type === 'idle') {
+        return true;
+      }
+      return false;
+    };
+    if (check(topType, topProps)) {
+      return true;
+    }
+    const payload = event.payload;
+    if (payload) {
+      return check(
+        payload.type || topType,
+        payload.properties || topProps
+      );
+    }
+    return false;
+  }
+
+  // ---------- Private streaming helpers ----------
+
+  private _isSystemEvent(type: string): boolean {
+    return (
+      type.startsWith('server.') ||
+      type.startsWith('workspace.') ||
+      type.startsWith('worktree.') ||
+      type.startsWith('lsp.') ||
+      type.startsWith('mcp.') ||
+      type.startsWith('pty.') ||
+      type.startsWith('installation.') ||
+      type === 'global.disposed' ||
+      // Session-level control events are not part of the assistant reply
+      // and would be confusing as separate lines in the history.
+      type === 'session.next.agent.switched' ||
+      type === 'session.next.model.switched' ||
+      type === 'file.edited'
+    );
+  }
+
+  private _appendToAssistantText(text: string): void {
+    if (!this._currentAssistantEl) {
+      this._currentAssistantEl = this._createMessageElement('assistant', '');
+    }
+    // Accumulate the raw markdown source, then re-render the whole
+    // message through marked.parse. This way the user sees properly
+    // formatted output (code blocks as <pre><code>, headings, lists)
+    // in real-time as text deltas arrive — NOT the raw ```fence```
+    // source that an append-only .textContent would show.
+    this._currentAssistantText += text;
+    this._renderAssistantMarkdown(
+      this._currentAssistantEl,
+      this._currentAssistantText
+    );
+    this._scrollToBottom();
+  }
+
+  private _getOrCreateBlock(key: string, title: string): BlockEntry {
+    let entry = this._activeBlocks[key];
+    if (entry) {
+      return entry;
+    }
+    const details = document.createElement('details');
+    details.className = `opencode-msg-block opencode-msg-block-${key}`;
+    details.open = true;
+    const summary = document.createElement('summary');
+    summary.textContent = title;
+    const content = document.createElement('div');
+    content.className = 'opencode-msg-block-content';
+    details.appendChild(summary);
+    details.appendChild(content);
+    this._historyEl.appendChild(details);
+    entry = { details, content };
+    this._activeBlocks[key] = entry;
+    this._scrollToBottom();
+    return entry;
+  }
+
+  private _appendToBlock(
+    key: string,
+    defaultTitle: string,
+    text: string
+  ): void {
+    const block = this._getOrCreateBlock(key, defaultTitle);
+    block.content.textContent = (block.content.textContent || '') + text;
+    this._scrollToBottom();
+  }
+
+  private _createPermissionBlock(permId: string, description: string): void {
+    if (!permId) {
+      return;
+    }
+    const domId = `opencode-perm-${permId}`;
+    if (this._historyEl.querySelector('#' + domId)) {
+      return;
+    }
+    const details = document.createElement('details');
+    details.id = domId;
+    details.className =
+      'opencode-msg-block opencode-msg-block-interaction';
+    details.open = true;
+    const summary = document.createElement('summary');
+    summary.textContent = '🔐 权限请求等待处理';
+    const content = document.createElement('div');
+    content.className = 'opencode-msg-block-content opencode-perm-body';
+
+    const desc = document.createElement('div');
+    desc.className = 'opencode-perm-desc';
+    desc.textContent = `请求内容：${description}`;
+    content.appendChild(desc);
+
+    const btnGroup = document.createElement('div');
+    btnGroup.className = 'opencode-perm-buttons';
+    const mkBtn = (
+      label: string,
+      cls: string,
+      resp: 'once' | 'always' | 'reject'
+    ): HTMLButtonElement => {
+      const b = document.createElement('button');
+      b.type = 'button';
+      b.className = cls;
+      b.textContent = label;
+      b.addEventListener('click', () => {
+        this._handlePermissionResponse(permId, resp, btnGroup, details);
+      });
+      return b;
+    };
+    btnGroup.appendChild(
+      mkBtn('允许一次 (Once)', 'opencode-perm-btn-allow', 'once')
+    );
+    btnGroup.appendChild(
+      mkBtn('总是允许 (Always)', 'opencode-perm-btn-allow', 'always')
+    );
+    btnGroup.appendChild(
+      mkBtn('拒绝 (Reject)', 'opencode-perm-btn-reject', 'reject')
+    );
+    content.appendChild(btnGroup);
+
+    details.appendChild(summary);
+    details.appendChild(content);
+    this._historyEl.appendChild(details);
+    this._scrollToBottom();
+  }
+
+  private async _handlePermissionResponse(
+    permId: string,
+    response: 'once' | 'always' | 'reject',
+    btnGroup: HTMLElement,
+    details: HTMLElement
+  ): Promise<void> {
+    if (!this._options.onPermissionReply) {
+      btnGroup.textContent = '(未配置 onPermissionReply 回调)';
+      return;
+    }
+    btnGroup.textContent = '正在提交…';
+    try {
+      const ok = await this._options.onPermissionReply(permId, response);
+      const labels: Record<string, string> = {
+        once: '✓ 已授权 (仅一次)',
+        always: '✓ 已授权 (总是允许)',
+        reject: '✕ 已拒绝操作'
+      };
+      const color = response === 'reject' ? '#ef4444' : '#10b981';
+      btnGroup.innerHTML = `<span style="color: ${color}; font-weight: bold;">${
+        labels[response]
+      }${ok ? '' : ' (后端失败)'}</span>`;
+    } catch (e) {
+      btnGroup.innerHTML = `<span style="color: #ef4444; font-size: 11px;">提交失败: ${(e as Error).message}</span>`;
+    }
+  }
+
+  private _createQuestionBlock(questionId: string, questionText: string): void {
+    if (!questionId) {
+      return;
+    }
+    const domId = `opencode-q-${questionId}`;
+    if (this._historyEl.querySelector('#' + domId)) {
+      return;
+    }
+    const details = document.createElement('details');
+    details.id = domId;
+    details.className =
+      'opencode-msg-block opencode-msg-block-interaction';
+    details.open = true;
+    const summary = document.createElement('summary');
+    summary.textContent = '❓ Agent 提问等待回复';
+    const content = document.createElement('div');
+    content.className = 'opencode-msg-block-content opencode-q-body';
+
+    const desc = document.createElement('div');
+    desc.className = 'opencode-q-desc';
+    desc.textContent = `问题：${questionText}`;
+    content.appendChild(desc);
+
+    const inputRow = document.createElement('div');
+    inputRow.className = 'opencode-q-input-row';
+    const input = document.createElement('input');
+    input.type = 'text';
+    input.className = 'opencode-q-input';
+    input.placeholder = '请输入你的回答…';
+    const submit = document.createElement('button');
+    submit.type = 'button';
+    submit.className = 'opencode-q-submit';
+    submit.textContent = '提交回答';
+    submit.addEventListener('click', () => {
+      void this._handleQuestionResponse(
+        questionId,
+        input,
+        inputRow,
+        details
+      );
+    });
+    input.addEventListener('keydown', (e) => {
+      if (e.key === 'Enter') {
+        e.preventDefault();
+        submit.click();
+      }
+    });
+    inputRow.appendChild(input);
+    inputRow.appendChild(submit);
+    content.appendChild(inputRow);
+
+    details.appendChild(summary);
+    details.appendChild(content);
+    this._historyEl.appendChild(details);
+    this._scrollToBottom();
+    input.focus();
+  }
+
+  private async _handleQuestionResponse(
+    questionId: string,
+    input: HTMLInputElement,
+    inputRow: HTMLElement,
+    _details: HTMLElement
+  ): Promise<void> {
+    const answer = input.value.trim();
+    if (!answer) {
+      return;
+    }
+    if (!this._options.onQuestionReply) {
+      inputRow.innerHTML = '<span style="color: #64748b;">(未配置 onQuestionReply 回调)</span>';
+      return;
+    }
+    inputRow.innerHTML = '<span style="color: #64748b;">正在提交…</span>';
+    try {
+      const ok = await this._options.onQuestionReply(questionId, answer);
+      const color = ok ? '#10b981' : '#ef4444';
+      const text = ok ? `✓ 已回答：${answer}` : `提交失败 (后端 ok=false)`;
+      inputRow.innerHTML = `<span style="color: ${color}; font-weight: bold;">${text}</span>`;
+    } catch (e) {
+      inputRow.innerHTML = `<span style="color: #ef4444;">提交失败: ${(e as Error).message}</span>`;
+    }
+  }
+
+  private _resetStreamPointers(): void {
+    this._currentAssistantEl = null;
+    this._currentAssistantText = '';
+    this._activeBlocks = {};
+  }
+
+  // ---------- Static-message rendering (used by setMessages + streaming) ----------
+
+  private _renderAssistantMarkdown(wrap: HTMLElement, content: string): void {
+    // Idempotent: clear existing content first so this can be called
+    // repeatedly with the same `wrap` element (used by streaming —
+    // every text delta re-renders the whole message).
+    wrap.textContent = '';
+    const rendered = document.createElement('div');
+    rendered.className = 'opencode-msg-content';
+    rendered.innerHTML = marked.parse(content) as string;
+    const codeBlocks = rendered.querySelectorAll('pre');
+    codeBlocks.forEach(pre => {
+      const code = pre.querySelector('code');
+      if (!code) {
+        return;
+      }
+      const codeText = code.textContent ?? '';
+      const blockWrap = document.createElement('div');
+      blockWrap.className = 'opencode-code-block';
+      const toolbar = document.createElement('div');
+      toolbar.className = 'opencode-code-toolbar';
+      const mkBtn = (
+        label: string,
+        title: string,
+        onClick: () => void
+      ): HTMLButtonElement => {
+        const b = document.createElement('button');
+        b.type = 'button';
+        b.className = 'opencode-code-btn';
+        b.textContent = label;
+        b.title = title;
+        b.addEventListener('click', onClick);
+        return b;
+      };
+      toolbar.appendChild(
+        mkBtn('📋 复制', '复制代码', () => {
+          void this._copyToClipboard(codeText);
+        })
+      );
+      toolbar.appendChild(
+        mkBtn('↪ 插入', '在光标位置插入', () => {
+          this._insertAtCursor(codeText);
+        })
+      );
+      toolbar.appendChild(
+        mkBtn('⟳ 替换', '替换单元格内容', () => {
+          this._replaceCell(codeText);
+        })
+      );
+      blockWrap.appendChild(toolbar);
+      pre.parentNode!.insertBefore(blockWrap, pre);
+      blockWrap.appendChild(pre);
+    });
+    wrap.appendChild(rendered);
+  }
+
+  private _createMessageElement(role: string, text: string): HTMLDivElement {
+    const div = document.createElement('div');
+    div.className = `opencode-msg opencode-msg-${role}`;
+    div.textContent = text;
+    this._historyEl.appendChild(div);
+    this._scrollToBottom();
+    return div;
+  }
+
+  private _scrollToBottom(): void {
+    this._historyEl.scrollTop = this._historyEl.scrollHeight;
+  }
+
+  // ---------- Cell-edit actions (Copy / Insert / Replace) ----------
+
   private async _copyToClipboard(text: string): Promise<void> {
     try {
       if (
@@ -335,11 +908,6 @@ export class OpenCodeInlinePrompt extends Widget {
     }
   }
 
-  /**
-   * Insert text at the current editor cursor position. If the cell has
-   * no editor (e.g. a non-Code cell) or the cursor can't be read, falls
-   * back to appending at the end of the cell.
-   */
   private _insertAtCursor(text: string): void {
     const cell = this._cell;
     if (!cell || !cell.model || !cell.model.sharedModel) {
@@ -367,7 +935,6 @@ export class OpenCodeInlinePrompt extends Widget {
         offset += col;
       }
     } catch {
-      // Fall back to append on any error reading the cursor.
       offset = source.length;
     }
     const newSource = source.slice(0, offset) + text + source.slice(offset);
@@ -375,7 +942,6 @@ export class OpenCodeInlinePrompt extends Widget {
     Notification.info('已插入到光标位置');
   }
 
-  /** Replace the entire cell source with the given text. */
   private _replaceCell(text: string): void {
     const cell = this._cell;
     if (!cell || !cell.model || !cell.model.sharedModel) {

src/types.ts

@@ -1,12 +1,13 @@
 /**
  * Shared types for the opencode-bridge frontend.
  *
- * These mirror the Python server's `CellContext` / `OpenCodeRequest` / `OpenCodeResponse`
- * shapes in `opencode_bridge/routes.py`. Keep them in sync.
+ * Mirrors the Python server's response shapes in `opencode_bridge/routes.py`
+ * and the OpenCode Serve SSE event shape documented in
+ * `/Users/taochen/temp/demo.html` (`handleGlobalEvent` / `processSessionEvent`).
  *
- * v3-final: the JupyterLab plugin settings have been removed entirely.
- * All connection config lives in startup environment variables (see
- * `opencode_bridge/config.py`); the inline model picker is fully dynamic.
+ * v4: POST /edit is async (returns immediately with sessionId). The LLM
+ * reply is consumed via the GET /events SSE stream. Each SSE event is
+ * a JSON object with a `type` string; consumers route on `type`.
  */
 
 export interface ErrorOutput {
@@ -33,21 +34,41 @@ export interface OpenCodeRequest {
   modelId?: string;
 }
 
-export interface OpenCodeSuccess {
+/**
+ * Response from POST /opencode-bridge/edit (async).
+ * The actual assistant reply arrives via the /events SSE stream and is
+ * NOT embedded in this response.
+ */
+export interface OpenCodeEditSuccess {
   ok: true;
-  /** The AI's reply as markdown (code in ```fences```, optional explanation).
-   *  Render with marked. The cell source is NOT replaced. */
-  markdown: string;
   sessionId: string;
   notebookPath: string;
 }
 
-export interface OpenCodeFailure {
+export interface OpenCodeEditFailure {
   ok: false;
   error: string;
 }
 
-export type OpenCodeResponse = OpenCodeSuccess | OpenCodeFailure;
+export type OpenCodeEditResponse = OpenCodeEditSuccess | OpenCodeEditFailure;
+
+/**
+ * One SSE event from GET /opencode-bridge/events (proxied from OpenCode
+ * Serve's /global/event). The shape is loosely typed because OpenCode's
+ * payload envelope is not strictly documented; we follow demo.html's
+ * defensive property lookup pattern (try top-level then nested under
+ * `payload`).
+ */
+export interface OpenCodeEvent {
+  type: string;
+  payload?: {
+    type?: string;
+    properties?: { [k: string]: any };
+    syncEvent?: { aggregateID?: string };
+  };
+  properties?: { [k: string]: any };
+  syncEvent?: { aggregateID?: string };
+}
 
 /** Response from GET /opencode-bridge/session-messages?notebook=... .
  *  Projected from OpenCode's {info, parts}[] by the server into a

style/base.css

@@ -210,3 +210,132 @@
   opacity: 0.4;
   cursor: default;
 }
+
+/* ------------------------------------------------------------------
+   Streaming event blocks (v4+ — see opencode_inline_prompt.applyEvent)
+   ------------------------------------------------------------------ */
+
+/* System / status line rendered for server.*, workspace.*, lsp.*, etc. */
+.opencode-inline-prompt .opencode-msg-system {
+  font-size: var(--jp-ui-font-size0, 11px);
+  color: var(--jp-ui-font-color2, #888);
+  padding: 2px 4px;
+  font-family: var(--jp-code-font-family, monospace);
+  white-space: pre-wrap;
+  word-break: break-all;
+}
+
+/* Base collapsible block used for reasoning / tool / permission / question. */
+.opencode-inline-prompt .opencode-msg-block {
+  margin: 4px 0;
+  border: 1px solid var(--jp-border-color2, #ddd);
+  border-radius: 3px;
+  background: var(--jp-layout-color1, #fafafa);
+  font-size: var(--jp-ui-font-size0, 11px);
+  overflow: hidden;
+}
+
+.opencode-inline-prompt .opencode-msg-block > summary {
+  padding: 2px 6px;
+  font-weight: 600;
+  cursor: pointer;
+  background: var(--jp-layout-color2, #f1f5f9);
+  color: var(--jp-ui-font-color2, #64748b);
+  user-select: none;
+  font-size: var(--jp-ui-font-size0, 11px);
+}
+
+.opencode-inline-prompt .opencode-msg-block-content {
+  padding: 4px 8px;
+  font-family: var(--jp-code-font-family, monospace);
+  white-space: pre-wrap;
+  word-break: break-all;
+  color: var(--jp-ui-font-color1, #334155);
+  max-height: 240px;
+  overflow-y: auto;
+}
+
+/* Reasoning: green tint (matches demo.html). */
+.opencode-inline-prompt .opencode-msg-block-reasoning > summary {
+  background: #f0fdf4;
+  color: #15803d;
+}
+
+/* Tool call: yellow tint. */
+.opencode-inline-prompt .opencode-msg-block-tool > summary {
+  background: #fef3c7;
+  color: #b45309;
+}
+
+/* Permission / question: pink tint. */
+.opencode-inline-prompt .opencode-msg-block-interaction > summary {
+  background: #fce7f3;
+  color: #be185d;
+}
+
+/* Permission block: description + 3 buttons. */
+.opencode-inline-prompt .opencode-perm-desc {
+  margin-bottom: 6px;
+  font-size: var(--jp-ui-font-size0, 11px);
+  color: var(--jp-ui-font-color1, #334155);
+}
+.opencode-inline-prompt .opencode-perm-buttons {
+  display: flex;
+  gap: 6px;
+  flex-wrap: wrap;
+  margin-top: 4px;
+}
+.opencode-inline-prompt .opencode-perm-btn-allow,
+.opencode-inline-prompt .opencode-perm-btn-reject {
+  border: none;
+  padding: 4px 10px;
+  border-radius: 3px;
+  cursor: pointer;
+  font-size: var(--jp-ui-font-size0, 11px);
+  color: white;
+  font-weight: 500;
+}
+.opencode-inline-prompt .opencode-perm-btn-allow {
+  background: #10b981;
+}
+.opencode-inline-prompt .opencode-perm-btn-allow:hover {
+  background: #059669;
+}
+.opencode-inline-prompt .opencode-perm-btn-reject {
+  background: #ef4444;
+}
+.opencode-inline-prompt .opencode-perm-btn-reject:hover {
+  background: #dc2626;
+}
+
+/* Question block: text + input + submit. */
+.opencode-inline-prompt .opencode-q-desc {
+  margin-bottom: 6px;
+  font-size: var(--jp-ui-font-size0, 11px);
+  color: var(--jp-ui-font-color1, #334155);
+}
+.opencode-inline-prompt .opencode-q-input-row {
+  display: flex;
+  gap: 6px;
+  margin-top: 4px;
+}
+.opencode-inline-prompt .opencode-q-input {
+  flex: 1;
+  padding: 3px 6px;
+  border: 1px solid var(--jp-border-color2, #ccc);
+  border-radius: 3px;
+  font-size: var(--jp-ui-font-size0, 11px);
+  font-family: inherit;
+  background: var(--jp-layout-color1, #fff);
+}
+.opencode-inline-prompt .opencode-q-submit {
+  border: 1px solid var(--jp-border-color2, #ccc);
+  background: var(--jp-layout-color2, #f0f0f0);
+  padding: 3px 10px;
+  border-radius: 3px;
+  cursor: pointer;
+  font-size: var(--jp-ui-font-size0, 11px);
+}
+.opencode-inline-prompt .opencode-q-submit:hover {
+  background: var(--jp-layout-color3, #e0e0e0);
+}